Merge pull request #57 from BobbyL2k/feat/llama-cpp-proxy

feat: Proxy llama.cpp API endpoints via `/llama-cpp/{name}/`

.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
+*~

README.md

@@ -23,6 +23,7 @@
 ### ⚡ Smart Operations
 - **Instance Monitoring**: Health checks, auto-restart, log management
 - **Smart Resource Management**: Idle timeout, LRU eviction, and configurable instance limits
+- **Environment Variables**: Set custom environment variables per instance for advanced configuration  
 
 ![Dashboard Screenshot](docs/images/dashboard.png)
 
@@ -52,7 +53,8 @@ llamactl
 2. Click "Create Instance"
 3. Choose backend type (llama.cpp, MLX, or vLLM)
 4. Set model path and backend-specific options
-5. Start or stop the instance
+5. Configure environment variables if needed (optional)
+6. Start or stop the instance
 
 ### Or use the REST API:
 ```bash
@@ -66,10 +68,10 @@ curl -X POST localhost:8080/api/v1/instances/my-mlx-model \
   -H "Authorization: Bearer your-key" \
   -d '{"backend_type": "mlx_lm", "backend_options": {"model": "mlx-community/Mistral-7B-Instruct-v0.3-4bit"}}'
 
-# Create vLLM instance
+# Create vLLM instance with environment variables
 curl -X POST localhost:8080/api/v1/instances/my-vllm-model \
   -H "Authorization: Bearer your-key" \
-  -d '{"backend_type": "vllm", "backend_options": {"model": "microsoft/DialoGPT-medium", "tensor_parallel_size": 2}}'
+  -d '{"backend_type": "vllm", "backend_options": {"model": "microsoft/DialoGPT-medium", "tensor_parallel_size": 2}, "environment": {"CUDA_VISIBLE_DEVICES": "0,1", "NCCL_DEBUG": "INFO"}}'
 
 # Use with OpenAI SDK
 curl -X POST localhost:8080/v1/chat/completions \
@@ -93,7 +95,30 @@ 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 .
+```
+
+**Features:** CUDA support, automatic latest release installation, no backend dependencies.
+**Note:** Dockerfiles are configured for CUDA. Adapt base images for other platforms (CPU, ROCm, etc.).
+
+For detailed Docker setup and configuration, see the [Installation Guide](docs/getting-started/installation.md).
+
+### Option 3: Build from Source
 Requires Go 1.24+ and Node.js 22+
 ```bash
 git clone https://github.com/lordmathis/llamactl.git
@@ -145,47 +170,23 @@ pip install vllm
 # Or use Docker - no local installation required
 ```
 
-## Docker Support
+## Backend Docker Support
 
-llamactl supports running backends in Docker containers with identical behavior to native execution. This is particularly useful for:
+llamactl can run backends in Docker containers:
-- Production deployments without local backend installation
-- Isolating backend dependencies
-- GPU-accelerated inference using official Docker images
-
-### Docker Configuration
-
-Enable Docker support using the new structured backend configuration:
 
 ```yaml
 backends:
   llama-cpp:
-    command: "llama-server"
     docker:
       enabled: true
-      image: "ghcr.io/ggml-org/llama.cpp:server"
-      args: ["run", "--rm", "--network", "host", "--gpus", "all"]
-
   vllm:
-    command: "vllm"
-    args: ["serve"]
     docker:
       enabled: true
-      image: "vllm/vllm-openai:latest"
-      args: ["run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g"]
 ```
 
-### Key Features
+**Requirements:** Docker installed and running. For GPU support: nvidia-docker2 (Linux) or Docker Desktop with GPU support.
 
-- **Host Networking**: Uses `--network host` for seamless port management
+For detailed Docker configuration options, see the [Configuration Guide](docs/getting-started/configuration.md).
-- **GPU Support**: Includes `--gpus all` for GPU acceleration
-- **Environment Variables**: Configure container environment as needed
-- **Flexible Configuration**: Per-backend Docker settings with sensible defaults
-
-### Requirements
-
-- Docker installed and running
-- For GPU support: nvidia-docker2 (Linux) or Docker Desktop with GPU support
-- No local backend installation required when using Docker
 
 ## Configuration
 
@@ -196,30 +197,34 @@ 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: {}
+      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"]
-      environment: {}
+      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

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 apidocs/ ./apidocs/
+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/getting-started/configuration.md

@@ -17,30 +17,37 @@ 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
@@ -98,6 +105,7 @@ 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)
 ```
 
@@ -113,36 +121,74 @@ backends:
   llama-cpp:
     command: "llama-server"
     args: []
+    environment: {}              # Environment variables for the backend process
     docker:
-      enabled: false                   # Enable Docker runtime (default: false)
+      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
+      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

docs/getting-started/installation.md

@@ -71,7 +71,72 @@ sudo mv llamactl /usr/local/bin/
 # Windows - Download from releases page
 ```
 
-### Option 2: Build from Source
+### 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
+
+**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
+```
+
+**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
+```
+
+**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

docs/user-guide/api-reference.md

@@ -116,7 +116,18 @@ Create and start a new instance.
 POST /api/v1/instances/{name}
 ```
 
-**Request Body:** JSON object with instance configuration. See [Managing Instances](managing-instances.md) for available configuration options.
+**Request Body:** JSON object with instance configuration. Common fields include:
+
+- `backend_type`: Backend type (`llama_cpp`, `mlx_lm`, or `vllm`)
+- `backend_options`: Backend-specific configuration
+- `auto_restart`: Enable automatic restart on failure
+- `max_restarts`: Maximum restart attempts
+- `restart_delay`: Delay between restarts in seconds
+- `on_demand_start`: Start instance when receiving requests
+- `idle_timeout`: Idle timeout in minutes
+- `environment`: Environment variables as key-value pairs
+
+See [Managing Instances](managing-instances.md) for complete configuration options.
 
 **Response:**
 ```json
@@ -354,7 +365,15 @@ 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"
+    "backend_type": "llama_cpp",
+    "backend_options": {
+      "model": "/models/llama-2-7b.gguf",
+      "gpu_layers": 32
+    },
+    "environment": {
+      "CUDA_VISIBLE_DEVICES": "0",
+      "OMP_NUM_THREADS": "8"
+    }
   }'
 
 # Check instance status

docs/user-guide/managing-instances.md

@@ -53,6 +53,7 @@ Each instance is displayed as a card showing:
     - **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
 6. Configure backend-specific options:
     - **llama.cpp**: Threads, context size, GPU layers, port, etc.
     - **MLX**: Temperature, top-p, adapter path, Python environment, etc.
@@ -101,7 +102,12 @@ curl -X POST http://localhost:8080/api/instances/my-vllm-instance \
       "gpu_memory_utilization": 0.9
     },
     "auto_restart": true,
-    "on_demand_start": true
+    "on_demand_start": true,
+    "environment": {
+      "CUDA_VISIBLE_DEVICES": "0,1",
+      "NCCL_DEBUG": "INFO",
+      "PYTHONPATH": "/custom/path"
+    }
   }'
 
 # Create llama.cpp instance with HuggingFace model

pkg/config/config.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"log"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -12,9 +13,11 @@ import (
 
 // BackendSettings contains structured backend configuration
 type BackendSettings struct {
-	Command string          `yaml:"command"`
+	Command         string            `yaml:"command"`
-	Args    []string        `yaml:"args"`
+	Args            []string          `yaml:"args"`
-	Docker  *DockerSettings `yaml:"docker,omitempty"`
+	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
@@ -54,8 +57,14 @@ 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
@@ -130,12 +139,14 @@ 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,
 		},
 		Backends: BackendConfig{
 			LlamaCpp: BackendSettings{
-				Command: "llama-server",
+				Command:     "llama-server",
-				Args:    []string{},
+				Args:        []string{},
+				Environment: map[string]string{},
 				Docker: &DockerSettings{
 					Enabled: false,
 					Image:   "ghcr.io/ggml-org/llama.cpp:server",
@@ -165,10 +176,12 @@ func LoadConfig(configPath string) (AppConfig, error) {
 			},
 		},
 		Instances: InstancesConfig{
-			PortRange:            [2]int{8000, 9000},
+			PortRange: [2]int{8000, 9000},
-			DataDir:              getDefaultDataDirectory(),
+			DataDir:   getDefaultDataDirectory(),
-			InstancesDir:         filepath.Join(getDefaultDataDirectory(), "instances"),
+			// NOTE: empty strings are set as placeholder values since InstancesDir and LogsDir
-			LogsDir:              filepath.Join(getDefaultDataDirectory(), "logs"),
+			// should be relative path to DataDir if not explicitly set.
+			InstancesDir:         "",
+			LogsDir:              "",
 			AutoCreateDirs:       true,
 			MaxInstances:         -1, // -1 means unlimited
 			MaxRunningInstances:  -1, // -1 means unlimited
@@ -196,6 +209,14 @@ func LoadConfig(configPath string) (AppConfig, error) {
 	// 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
 }
 
@@ -216,6 +237,7 @@ 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
 		}
 	}
@@ -288,6 +310,12 @@ func loadEnvVars(cfg *AppConfig) {
 	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 {
@@ -315,18 +343,28 @@ func loadEnvVars(cfg *AppConfig) {
 		if cfg.Backends.LlamaCpp.Docker.Environment == nil {
 			cfg.Backends.LlamaCpp.Docker.Environment = make(map[string]string)
 		}
-		// Parse env vars in format "KEY1=value1,KEY2=value2"
+		parseEnvVars(llamaDockerEnv, cfg.Backends.LlamaCpp.Docker.Environment)
-		for _, envPair := range strings.Split(llamaDockerEnv, ",") {
+	}
-			if parts := strings.SplitN(strings.TrimSpace(envPair), "=", 2); len(parts) == 2 {
+	if llamaEnv := os.Getenv("LLAMACTL_LLAMACPP_RESPONSE_HEADERS"); llamaEnv != "" {
-				cfg.Backends.LlamaCpp.Docker.Environment[parts[0]] = parts[1]
+		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 {
@@ -354,12 +392,13 @@ func loadEnvVars(cfg *AppConfig) {
 		if cfg.Backends.VLLM.Docker.Environment == nil {
 			cfg.Backends.VLLM.Docker.Environment = make(map[string]string)
 		}
-		// Parse env vars in format "KEY1=value1,KEY2=value2"
+		parseEnvVars(vllmDockerEnv, cfg.Backends.VLLM.Docker.Environment)
-		for _, envPair := range strings.Split(vllmDockerEnv, ",") {
+	}
-			if parts := strings.SplitN(strings.TrimSpace(envPair), "=", 2); len(parts) == 2 {
+	if llamaEnv := os.Getenv("LLAMACTL_VLLM_RESPONSE_HEADERS"); llamaEnv != "" {
-				cfg.Backends.VLLM.Docker.Environment[parts[0]] = parts[1]
+		if cfg.Backends.VLLM.ResponseHeaders == nil {
-			}
+			cfg.Backends.VLLM.ResponseHeaders = make(map[string]string)
 		}
+		parseHeaders(llamaEnv, cfg.Backends.VLLM.ResponseHeaders)
 	}
 
 	// MLX backend
@@ -369,6 +408,18 @@ func loadEnvVars(cfg *AppConfig) {
 	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 != "" {
@@ -443,6 +494,32 @@ 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 {
@@ -475,6 +552,10 @@ 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/instance/instance.go

@@ -198,6 +198,15 @@ func (i *Process) GetProxy() (*httputil.ReverseProxy, error) {
 
 	proxy := httputil.NewSingleHostReverseProxy(targetURL)
 
+	var responseHeaders map[string]string
+	switch i.options.BackendType {
+	case backends.BackendTypeLlamaCpp:
+		responseHeaders = i.globalBackendSettings.LlamaCpp.ResponseHeaders
+	case backends.BackendTypeVllm:
+		responseHeaders = i.globalBackendSettings.VLLM.ResponseHeaders
+	case backends.BackendTypeMlxLm:
+		responseHeaders = i.globalBackendSettings.MLX.ResponseHeaders
+	}
 	proxy.ModifyResponse = func(resp *http.Response) error {
 		// Remove CORS headers from llama-server response to avoid conflicts
 		// llamactl will add its own CORS headers
@@ -207,6 +216,10 @@ func (i *Process) GetProxy() (*httputil.ReverseProxy, error) {
 		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 responseHeaders {
+			resp.Header.Set(key, value)
+		}
 		return nil
 	}
 

pkg/instance/lifecycle.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"log"
 	"net/http"
+	"os"
 	"os/exec"
 	"runtime"
 	"syscall"
@@ -37,6 +38,9 @@ func (i *Process) Start() error {
 	// Initialize last request time to current time when starting
 	i.lastRequestTime.Store(i.timeProvider.Now().Unix())
 
+	// Create context before building command (needed for CommandContext)
+	i.ctx, i.cancel = context.WithCancel(context.Background())
+
 	// Create log files
 	if err := i.logger.Create(); err != nil {
 		return fmt.Errorf("failed to create log files: %w", err)
@@ -47,8 +51,6 @@ func (i *Process) Start() error {
 	if cmdErr != nil {
 		return fmt.Errorf("failed to build command: %w", cmdErr)
 	}
-
-	i.ctx, i.cancel = context.WithCancel(context.Background())
 	i.cmd = cmd
 
 	if runtime.GOOS != "windows" {
@@ -372,13 +374,27 @@ func (i *Process) buildCommand() (*exec.Cmd, error) {
 		return nil, err
 	}
 
+	// Build the environment variables
+	env := i.options.BuildEnvironment(backendConfig)
+
 	// Get the command to execute
-	cmd := i.options.GetCommand(backendConfig)
+	command := i.options.GetCommand(backendConfig)
 
 	// Build command arguments
 	args := i.options.BuildCommandArgs(backendConfig)
 
-	return exec.Command(cmd, args...), nil
+	// Create the exec.Cmd
+	cmd := exec.CommandContext(i.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
 }
 
 // getBackendConfig resolves the backend configuration for the current instance

pkg/instance/options.go

@@ -9,6 +9,7 @@ import (
 	"llamactl/pkg/backends/vllm"
 	"llamactl/pkg/config"
 	"log"
+	"maps"
 )
 
 type CreateInstanceOptions struct {
@@ -20,6 +21,8 @@ type CreateInstanceOptions struct {
 	OnDemandStart *bool `json:"on_demand_start,omitempty"`
 	// Idle timeout
 	IdleTimeout *int `json:"idle_timeout,omitempty"` // minutes
+	//Environment variables
+	Environment map[string]string `json:"environment,omitempty"`
 
 	BackendType    backends.BackendType `json:"backend_type"`
 	BackendOptions map[string]any       `json:"backend_options,omitempty"`
@@ -240,3 +243,23 @@ func (c *CreateInstanceOptions) BuildCommandArgs(backendConfig *config.BackendSe
 
 	return args
 }
+
+func (c *CreateInstanceOptions) BuildEnvironment(backendConfig *config.BackendSettings) map[string]string {
+	env := map[string]string{}
+
+	if backendConfig.Environment != nil {
+		maps.Copy(env, backendConfig.Environment)
+	}
+
+	if backendConfig.Docker != nil && backendConfig.Docker.Enabled && c.BackendType != backends.BackendTypeMlxLm {
+		if backendConfig.Docker.Environment != nil {
+			maps.Copy(env, backendConfig.Docker.Environment)
+		}
+	}
+
+	if c.Environment != nil {
+		maps.Copy(env, c.Environment)
+	}
+
+	return env
+}

pkg/manager/manager.go

@@ -263,19 +263,32 @@ func (im *instanceManager) loadInstance(name, path string) error {
 }
 
 // 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
+	var instancesToStop []*instance.Process
 	for _, inst := range im.instances {
 		if inst.IsRunning() && // Was running when persisted
 			inst.GetOptions() != nil &&
-			inst.GetOptions().AutoRestart != nil &&
+			inst.GetOptions().AutoRestart != nil {
-			*inst.GetOptions().AutoRestart {
+			if *inst.GetOptions().AutoRestart {
-			instancesToStart = append(instancesToStart, inst)
+				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)
+	}
+
+	// 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)

pkg/manager/manager_test.go

@@ -209,3 +209,66 @@ func createTestManager() manager.InstanceManager {
 	}
 	return manager.NewInstanceManager(backendConfig, cfg)
 }
+
+func TestAutoRestartDisabledInstanceStatus(t *testing.T) {
+	tempDir := t.TempDir()
+
+	backendConfig := config.BackendConfig{
+		LlamaCpp: config.BackendSettings{
+			Command: "llama-server",
+		},
+	}
+
+	cfg := config.InstancesConfig{
+		PortRange:            [2]int{8000, 9000},
+		InstancesDir:         tempDir,
+		MaxInstances:         10,
+		TimeoutCheckInterval: 5,
+	}
+
+	// Create first manager and instance with auto-restart disabled
+	manager1 := manager.NewInstanceManager(backendConfig, cfg)
+
+	autoRestart := false
+	options := &instance.CreateInstanceOptions{
+		BackendType: backends.BackendTypeLlamaCpp,
+		AutoRestart: &autoRestart,
+		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+			Model: "/path/to/model.gguf",
+			Port:  8080,
+		},
+	}
+
+	inst, err := manager1.CreateInstance("test-instance", options)
+	if err != nil {
+		t.Fatalf("CreateInstance failed: %v", err)
+	}
+
+	// Simulate instance being in running state when persisted
+	// (this would happen if the instance was running when llamactl was stopped)
+	inst.SetStatus(instance.Running)
+
+	// Shutdown first manager
+	manager1.Shutdown()
+
+	// Create second manager (simulating restart of llamactl)
+	manager2 := manager.NewInstanceManager(backendConfig, cfg)
+
+	// Get the loaded instance
+	loadedInst, err := manager2.GetInstance("test-instance")
+	if err != nil {
+		t.Fatalf("GetInstance failed: %v", err)
+	}
+
+	// The instance should be marked as Stopped, not Running
+	// because auto-restart is disabled
+	if loadedInst.IsRunning() {
+		t.Errorf("Expected instance with auto-restart disabled to be stopped after manager restart, but it was running")
+	}
+
+	if loadedInst.GetStatus() != instance.Stopped {
+		t.Errorf("Expected instance status to be Stopped, got %v", loadedInst.GetStatus())
+	}
+
+	manager2.Shutdown()
+}

pkg/server/handlers.go

@@ -131,11 +131,16 @@ func (h *Handler) ListInstances() http.HandlerFunc {
 			return
 		}
 
-		w.Header().Set("Content-Type", "application/json")
+		// Marshal to bytes first to set Content-Length header
-		if err := json.NewEncoder(w).Encode(instances); err != nil {
+		data, err := json.Marshal(instances)
+		if err != nil {
 			http.Error(w, "Failed to encode instances: "+err.Error(), http.StatusInternalServerError)
 			return
 		}
+
+		w.Header().Set("Content-Type", "application/json")
+		w.Header().Set("Content-Length", strconv.Itoa(len(data)))
+		w.Write(data)
 	}
 }
 
@@ -202,7 +207,7 @@ func (h *Handler) GetInstance() http.HandlerFunc {
 
 		inst, err := h.InstanceManager.GetInstance(name)
 		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
+			http.Error(w, "Invalid instance: "+err.Error(), http.StatusBadRequest)
 			return
 		}
 
@@ -475,29 +480,15 @@ func (h *Handler) ProxyToInstance() http.HandlerFunc {
 
 		// 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):]
+		r.URL.Path = strings.TrimPrefix(r.URL.Path, 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)
 	}
@@ -580,12 +571,13 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 		// Route to the appropriate inst based on instance name
 		inst, err := h.InstanceManager.GetInstance(modelName)
 		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
+			http.Error(w, "Invalid instance: "+err.Error(), http.StatusBadRequest)
 			return
 		}
 
 		if !inst.IsRunning() {
-			allowOnDemand := inst.GetOptions() != nil && inst.GetOptions().OnDemandStart != nil && *inst.GetOptions().OnDemandStart
+			options := inst.GetOptions()
+			allowOnDemand := options != nil && options.OnDemandStart != nil && *options.OnDemandStart
 			if !allowOnDemand {
 				http.Error(w, "Instance is not running", http.StatusServiceUnavailable)
 				return
@@ -634,6 +626,84 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 	}
 }
 
+func (h *Handler) LlamaCppProxy(onDemandStart bool) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+
+		// Get the instance name from the URL parameter
+		name := chi.URLParam(r, "name")
+		if name == "" {
+			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
+			return
+		}
+
+		// Route to the appropriate inst based on instance name
+		inst, err := h.InstanceManager.GetInstance(name)
+		if err != nil {
+			http.Error(w, "Invalid instance: "+err.Error(), http.StatusBadRequest)
+			return
+		}
+
+		options := inst.GetOptions()
+		if options == nil {
+			http.Error(w, "Cannot obtain Instance's options", http.StatusInternalServerError)
+			return
+		}
+
+		if options.BackendType != backends.BackendTypeLlamaCpp {
+			http.Error(w, "Instance is not a llama.cpp server.", http.StatusBadRequest)
+			return
+		}
+
+		if !inst.IsRunning() {
+
+			if !(onDemandStart && options.OnDemandStart != nil && *options.OnDemandStart) {
+				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(name); 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
+		}
+
+		// Strip the "/llama-cpp/<name>" prefix from the request URL
+		prefix := fmt.Sprintf("/llama-cpp/%s", name)
+		r.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
+
+		// Update the last request time for the instance
+		inst.UpdateLastRequestTime()
+
+		proxy.ServeHTTP(w, r)
+	}
+}
+
 // ParseCommandRequest represents the request body for command parsing
 type ParseCommandRequest struct {
 	Command string `json:"command"`

pkg/server/routes.go

@@ -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,
@@ -103,6 +103,51 @@ 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.LlamaCppProxy(false))
+
+		// 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(true)
+
+			// 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)

webui/src/components/ZodFormField.tsx

@@ -1,144 +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 { BackendType, type CreateInstanceOptions } 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}
-          </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>
-            <option value={BackendType.MLX_LM}>MLX LM</option>
-            <option value={BackendType.VLLM}>vLLM</option>
-          </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}
-              </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}
-              </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}
-              </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/form/EnvironmentVariablesInput.tsx

@@ -0,0 +1,144 @@
+import React, { useState } from 'react'
+import { Input } from '@/components/ui/input'
+import { Label } from '@/components/ui/label'
+import { Button } from '@/components/ui/button'
+import { X, Plus } from 'lucide-react'
+
+interface EnvironmentVariablesInputProps {
+  id: string
+  label: string
+  value: Record<string, string> | undefined
+  onChange: (value: Record<string, string> | undefined) => void
+  description?: string
+  disabled?: boolean
+  className?: string
+}
+
+interface EnvVar {
+  key: string
+  value: string
+}
+
+const EnvironmentVariablesInput: React.FC<EnvironmentVariablesInputProps> = ({
+  id,
+  label,
+  value,
+  onChange,
+  description,
+  disabled = false,
+  className
+}) => {
+  // Convert the value object to an array of key-value pairs for editing
+  const envVarsFromValue = value
+    ? Object.entries(value).map(([key, val]) => ({ key, value: val }))
+    : []
+
+  const [envVars, setEnvVars] = useState<EnvVar[]>(
+    envVarsFromValue.length > 0 ? envVarsFromValue : [{ key: '', value: '' }]
+  )
+
+  // Update parent component when env vars change
+  const updateParent = (newEnvVars: EnvVar[]) => {
+    // Filter out empty entries
+    const validVars = newEnvVars.filter(env => env.key.trim() !== '' && env.value.trim() !== '')
+
+    if (validVars.length === 0) {
+      onChange(undefined)
+    } else {
+      const envObject = validVars.reduce((acc, env) => {
+        acc[env.key.trim()] = env.value.trim()
+        return acc
+      }, {} as Record<string, string>)
+      onChange(envObject)
+    }
+  }
+
+  const handleKeyChange = (index: number, newKey: string) => {
+    const newEnvVars = [...envVars]
+    newEnvVars[index].key = newKey
+    setEnvVars(newEnvVars)
+    updateParent(newEnvVars)
+  }
+
+  const handleValueChange = (index: number, newValue: string) => {
+    const newEnvVars = [...envVars]
+    newEnvVars[index].value = newValue
+    setEnvVars(newEnvVars)
+    updateParent(newEnvVars)
+  }
+
+  const addEnvVar = () => {
+    const newEnvVars = [...envVars, { key: '', value: '' }]
+    setEnvVars(newEnvVars)
+  }
+
+  const removeEnvVar = (index: number) => {
+    if (envVars.length === 1) {
+      // Reset to empty if it's the last one
+      const newEnvVars = [{ key: '', value: '' }]
+      setEnvVars(newEnvVars)
+      updateParent(newEnvVars)
+    } else {
+      const newEnvVars = envVars.filter((_, i) => i !== index)
+      setEnvVars(newEnvVars)
+      updateParent(newEnvVars)
+    }
+  }
+
+  return (
+    <div className={`grid gap-2 ${className || ''}`}>
+      <Label htmlFor={id}>
+        {label}
+      </Label>
+      <div className="space-y-2">
+        {envVars.map((envVar, index) => (
+          <div key={index} className="flex gap-2 items-center">
+            <Input
+              placeholder="Variable name"
+              value={envVar.key}
+              onChange={(e) => handleKeyChange(index, e.target.value)}
+              disabled={disabled}
+              className="flex-1"
+            />
+            <Input
+              placeholder="Variable value"
+              value={envVar.value}
+              onChange={(e) => handleValueChange(index, e.target.value)}
+              disabled={disabled}
+              className="flex-1"
+            />
+            <Button
+              type="button"
+              variant="outline"
+              size="sm"
+              onClick={() => removeEnvVar(index)}
+              disabled={disabled}
+              className="shrink-0"
+            >
+              <X className="h-4 w-4" />
+            </Button>
+          </div>
+        ))}
+        <Button
+          type="button"
+          variant="outline"
+          size="sm"
+          onClick={addEnvVar}
+          disabled={disabled}
+          className="w-fit"
+        >
+          <Plus className="h-4 w-4 mr-2" />
+          Add Variable
+        </Button>
+      </div>
+      {description && (
+        <p className="text-sm text-muted-foreground">{description}</p>
+      )}
+      <p className="text-xs text-muted-foreground">
+        Environment variables that will be passed to the backend process
+      </p>
+    </div>
+  )
+}
+
+export default EnvironmentVariablesInput

webui/src/components/instance/BasicInstanceFields.tsx

@@ -1,99 +0,0 @@
-import React from 'react'
-import { BackendType, type CreateInstanceOptions } from '@/types/instance'
-import { getBasicFields, basicFieldsConfig } from '@/lib/zodFormUtils'
-import { getFieldType } from '@/schemas/instanceOptions'
-import TextInput from '@/components/form/TextInput'
-import NumberInput from '@/components/form/NumberInput'
-import CheckboxInput from '@/components/form/CheckboxInput'
-import SelectInput from '@/components/form/SelectInput'
-
-interface BasicInstanceFieldsProps {
-  formData: CreateInstanceOptions
-  onChange: (key: keyof CreateInstanceOptions, value: any) => void
-}
-
-const BasicInstanceFields: React.FC<BasicInstanceFieldsProps> = ({
-  formData,
-  onChange
-}) => {
-  const basicFields = getBasicFields()
-
-  const renderField = (fieldKey: keyof CreateInstanceOptions) => {
-    const config = basicFieldsConfig[fieldKey as string] || { label: fieldKey }
-    const fieldType = getFieldType(fieldKey)
-
-    // Special handling for backend_type field
-    if (fieldKey === 'backend_type') {
-      return (
-        <SelectInput
-          key={fieldKey}
-          id={fieldKey}
-          label={config.label}
-          value={formData[fieldKey] || BackendType.LLAMA_CPP}
-          onChange={(value) => onChange(fieldKey, value)}
-          options={[
-            { value: BackendType.LLAMA_CPP, label: 'Llama Server' },
-            { value: BackendType.MLX_LM, label: 'MLX LM' },
-            { value: BackendType.VLLM, label: 'vLLM' }
-          ]}
-          description={config.description}
-        />
-      )
-    }
-
-    // Render based on field type
-    switch (fieldType) {
-      case 'boolean':
-        return (
-          <CheckboxInput
-            key={fieldKey}
-            id={fieldKey}
-            label={config.label}
-            value={formData[fieldKey] as boolean | undefined}
-            onChange={(value) => onChange(fieldKey, value)}
-            description={config.description}
-          />
-        )
-
-      case 'number':
-        return (
-          <NumberInput
-            key={fieldKey}
-            id={fieldKey}
-            label={config.label}
-            value={formData[fieldKey] as number | undefined}
-            onChange={(value) => onChange(fieldKey, value)}
-            placeholder={config.placeholder}
-            description={config.description}
-          />
-        )
-
-      default:
-        return (
-          <TextInput
-            key={fieldKey}
-            id={fieldKey}
-            label={config.label}
-            value={formData[fieldKey] as string | number | undefined}
-            onChange={(value) => onChange(fieldKey, value)}
-            placeholder={config.placeholder}
-            description={config.description}
-          />
-        )
-    }
-  }
-
-  // Filter out auto restart fields and backend_options (handled separately)
-  const fieldsToRender = basicFields.filter(
-    fieldKey => !['auto_restart', 'max_restarts', 'restart_delay', 'backend_options'].includes(fieldKey as string)
-  )
-
-  return (
-    <div className="space-y-4">
-      <h3 className="text-lg font-medium">Basic Configuration</h3>
-      {fieldsToRender.map(renderField)}
-    </div>
-  )
-}
-
-export default BasicInstanceFields

webui/src/components/instance/InstanceSettingsCard.tsx

@@ -6,6 +6,7 @@ import { Input } from '@/components/ui/input'
 import AutoRestartConfiguration from '@/components/instance/AutoRestartConfiguration'
 import NumberInput from '@/components/form/NumberInput'
 import CheckboxInput from '@/components/form/CheckboxInput'
+import EnvironmentVariablesInput from '@/components/form/EnvironmentVariablesInput'
 
 interface InstanceSettingsCardProps {
   instanceName: string
@@ -75,6 +76,14 @@ const InstanceSettingsCard: React.FC<InstanceSettingsCardProps> = ({
             onChange={(value) => onChange('on_demand_start', value)}
             description="Start instance only when needed"
           />
+
+          <EnvironmentVariablesInput
+            id="environment"
+            label="Environment Variables"
+            value={formData.environment}
+            onChange={(value) => onChange('environment', value)}
+            description="Custom environment variables for the instance"
+          />
         </div>
       </CardContent>
     </Card>

webui/src/contexts/AuthContext.tsx

@@ -1,4 +1,4 @@
-import { type ReactNode, createContext, useContext, useState, useEffect, useCallback } from 'react'
+import { type ReactNode, createContext, useCallback, useContext, useEffect, useState } from 'react'
 
 interface AuthContextState {
   isAuthenticated: boolean
@@ -62,7 +62,7 @@ export const AuthProvider = ({ children }: AuthProviderProps) => {
   // Validate API key by making a test request
   const validateApiKey = async (key: string): Promise<boolean> => {
     try {
-      const response = await fetch('/api/v1/instances', {
+      const response = await fetch(document.baseURI + 'api/v1/instances', {
         headers: {
           'Authorization': `Bearer ${key}`,
           'Content-Type': 'application/json'

webui/src/lib/__tests__/api.test.ts

@@ -1,5 +1,5 @@
-import { describe, it, expect, vi, beforeEach } from 'vitest'
 import { instancesApi } from '@/lib/api'
+import { beforeEach, describe, expect, it, vi } from 'vitest'
 
 // Mock fetch globally
 const mockFetch = vi.fn()
@@ -11,11 +11,13 @@ describe('API Error Handling', () => {
   })
 
   it('converts HTTP errors to meaningful messages', async () => {
-    mockFetch.mockResolvedValue({
+    const mockResponse = {
       ok: false,
       status: 409,
-      text: () => Promise.resolve('Instance already exists')
+      text: () => Promise.resolve('Instance already exists'),
-    })
+      clone: function() { return this }
+    }
+    mockFetch.mockResolvedValue(mockResponse)
 
     await expect(instancesApi.create('existing', {}))
       .rejects
@@ -23,11 +25,13 @@ describe('API Error Handling', () => {
   })
 
   it('handles empty error responses gracefully', async () => {
-    mockFetch.mockResolvedValue({
+    const mockResponse = {
       ok: false,
       status: 500,
-      text: () => Promise.resolve('')
+      text: () => Promise.resolve(''),
-    })
+      clone: function() { return this }
+    }
+    mockFetch.mockResolvedValue(mockResponse)
 
     await expect(instancesApi.list())
       .rejects
@@ -53,7 +57,9 @@ describe('API Error Handling', () => {
     await instancesApi.getLogs('test-instance', 100)
 
     expect(mockFetch).toHaveBeenCalledWith(
-      '/api/v1/instances/test-instance/logs?lines=100',
+      expect.stringMatching(
+        /^https?:\/\/[^/]+\/api\/v1\/instances\/test-instance\/logs\?lines=100$/
+      ),
       expect.any(Object)
     )
   })

webui/src/lib/api.ts

@@ -1,7 +1,10 @@
 import type { CreateInstanceOptions, Instance } from "@/types/instance";
 import { handleApiError } from "./errorUtils";
 
-const API_BASE = "/api/v1";
+// Adding baseURI as a prefix to support being served behind a subpath
+// e.g. when llmamctl's `/` is served behind a reverse proxy at `/proxy/...`
+// the baseURI will be `/proxy/` and the API calls will be made to `/proxy/api/v1/<endpoint>`
+export const API_BASE = document.baseURI + "api/v1";
 
 // Base API call function with error handling
 async function apiCall<T>(
@@ -46,11 +49,8 @@ async function apiCall<T>(
     } else {
       // Handle empty responses for JSON endpoints
       const contentLength = response.headers.get('content-length');
-      if (contentLength === '0' || contentLength === null) {
+      if (contentLength === '0') {
-        const text = await response.text();
+        return {} as T; // Return empty object for empty JSON responses
-        if (text.trim() === '') {
-          return {} as T; // Return empty object for empty JSON responses
-        }
       }
       const data = await response.json() as T;
       return data;

webui/src/lib/errorUtils.ts

@@ -26,7 +26,8 @@ export async function handleApiError(response: Response): Promise<void> {
   }
 
   if (!response.ok) {
-    const errorMessage = await parseErrorResponse(response)
+    // Clone the response before reading to avoid consuming the body stream
+    const errorMessage = await parseErrorResponse(response.clone())
     throw new Error(errorMessage)
   }
 }

webui/src/lib/zodFormUtils.ts

@@ -1,12 +1,10 @@
 import {
-  type CreateInstanceOptions,
   type LlamaCppBackendOptions,
   type MlxBackendOptions,
   type VllmBackendOptions,
   LlamaCppBackendOptionsSchema,
   MlxBackendOptionsSchema,
   VllmBackendOptionsSchema,
-  getAllFieldKeys,
   getAllLlamaCppFieldKeys,
   getAllMlxFieldKeys,
   getAllVllmFieldKeys,
@@ -15,41 +13,6 @@ import {
   getVllmFieldType
 } from '@/schemas/instanceOptions'
 
-// Instance-level basic fields (not backend-specific)
-export const basicFieldsConfig: Record<string, {
-  label: string
-  description?: string
-  placeholder?: string
-}> = {
-  auto_restart: {
-    label: 'Auto Restart',
-    description: 'Automatically restart the instance on failure'
-  },
-  max_restarts: {
-    label: 'Max Restarts',
-    placeholder: '3',
-    description: 'Maximum number of restart attempts (0 = unlimited)'
-  },
-  restart_delay: {
-    label: 'Restart Delay (seconds)',
-    placeholder: '5',
-    description: 'Delay in seconds before attempting restart'
-  },
-  idle_timeout: {
-    label: 'Idle Timeout (minutes)',
-    placeholder: '60',
-    description: 'Time in minutes before instance is considered idle and stopped'
-  },
-  on_demand_start: {
-    label: 'On-Demand Start',
-    description: 'Start instance upon receiving OpenAI-compatible API request'
-  },
-  backend_type: {
-    label: 'Backend Type',
-    description: 'Type of backend to use for this instance'
-  }
-}
-
 // LlamaCpp backend-specific basic fields
 const basicLlamaCppFieldsConfig: Record<string, {
   label: string
@@ -152,18 +115,6 @@ const backendFieldGetters = {
   llama_cpp: getAllLlamaCppFieldKeys,
 } as const
 
-function isBasicField(key: keyof CreateInstanceOptions): boolean {
-  return key in basicFieldsConfig
-}
-
-export function getBasicFields(): (keyof CreateInstanceOptions)[] {
-  return Object.keys(basicFieldsConfig) as (keyof CreateInstanceOptions)[]
-}
-
-export function getAdvancedFields(): (keyof CreateInstanceOptions)[] {
-  return getAllFieldKeys().filter(key => !isBasicField(key))
-}
-
 export function getBasicBackendFields(backendType?: string): string[] {
   const normalizedType = (backendType || 'llama_cpp') as keyof typeof backendFieldConfigs
   const config = backendFieldConfigs[normalizedType] || basicLlamaCppFieldsConfig
@@ -222,5 +173,3 @@ export function getBackendFieldType(key: string): 'text' | 'number' | 'boolean'
   return 'text'
 }
 
-// Re-export the Zod-based functions
-export { getFieldType } from '@/schemas/instanceOptions'

webui/src/schemas/instanceOptions.ts

@@ -33,6 +33,9 @@ export const CreateInstanceOptionsSchema = z.object({
   idle_timeout: z.number().optional(),
   on_demand_start: z.boolean().optional(),
 
+  // Environment variables
+  environment: z.record(z.string(), z.string()).optional(),
+
   // Backend configuration
   backend_type: z.enum([BackendType.LLAMA_CPP, BackendType.MLX_LM, BackendType.VLLM]).optional(),
   backend_options: BackendOptionsSchema.optional(),
@@ -75,5 +78,6 @@ export function getFieldType(key: keyof CreateInstanceOptions): 'text' | 'number
   if (innerSchema instanceof z.ZodNumber) return 'number'
   if (innerSchema instanceof z.ZodArray) return 'array'
   if (innerSchema instanceof z.ZodObject) return 'object'
+  if (innerSchema instanceof z.ZodRecord) return 'object' // Handle ZodRecord as object
   return 'text' // ZodString and others default to text
 }

webui/vite.config.ts

@@ -21,4 +21,6 @@ export default defineConfig({
     setupFiles: ['./src/test/setup.ts'],
     css: true,
   },
+  // ensures relative asset paths to support being served behind a subpath
+  base: "./"
 })