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

@@ -95,7 +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
@@ -147,9 +170,9 @@ pip install vllm
 # Or use Docker - no local installation required
 ```
 
-## Docker Support
+## Backend Docker Support
 
-llamactl supports running backends in Docker containers - perfect for production deployments without local backend installation. Simply enable Docker in your configuration:
+llamactl can run backends in Docker containers:
 
 ```yaml
 backends:
@@ -174,6 +197,7 @@ 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:

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

@@ -17,33 +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
+    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
+    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
+    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
@@ -101,6 +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)
 ```
 
@@ -116,40 +121,46 @@ backends:
   llama-cpp:
     command: "llama-server"
     args: []
-    environment: {}                    # Environment variables for the backend process
+    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
+    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
+    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:**
@@ -160,6 +171,7 @@ backends:
 - `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
@@ -169,11 +181,13 @@ backends:
 - `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
 

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

pkg/config/config.go

@@ -13,10 +13,11 @@ import (
 
 // 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"`
+	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
@@ -56,8 +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
@@ -132,6 +139,7 @@ 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{
@@ -337,6 +345,12 @@ func loadEnvVars(cfg *AppConfig) {
 		}
 		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 != "" {
@@ -380,6 +394,12 @@ func loadEnvVars(cfg *AppConfig) {
 		}
 		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 != "" {
@@ -394,6 +414,12 @@ func loadEnvVars(cfg *AppConfig) {
 		}
 		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 != "" {
@@ -481,6 +507,19 @@ func parseEnvVars(envString string, envMap map[string]string) {
 	}
 }
 
+// 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 {

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" {
@@ -383,7 +385,11 @@ func (i *Process) buildCommand() (*exec.Cmd, error) {
 
 	// Create the exec.Cmd
 	cmd := exec.CommandContext(i.ctx, command, args...)
-	cmd.Env = []string{}
+
+	// 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))
 	}

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 {
-			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)
+	}
+
+	// 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")
-		if err := json.NewEncoder(w).Encode(instances); err != nil {
+		// Marshal to bytes first to set Content-Length header
+		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):]
-
-		// Ensure the proxy path starts with "/"
-		if !strings.HasPrefix(proxyPath, "/") {
-			proxyPath = "/" + proxyPath
-		}
+		r.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
 
 		// 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/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) {
-        const text = await response.text();
-        if (text.trim() === '') {
-          return {} as T; // Return empty object for empty JSON responses
-        }
+      if (contentLength === '0') {
+        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/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: "./"
 })