Marker MCP Server (Node.js)

A Node.js MCP (Model Context Protocol) server wrapper for Marker, enabling AI assistants like Claude to directly convert documents, extract tables, and perform OCR.

🚀 Quick Start with NPX

The easiest way to use Marker MCP server is with NPX (no installation required):

# Run directly with NPX
npx marker-mcp start

# Or with a remote Marker server
npx marker-mcp start --url http://your-marker-server:8000

# Generate Claude Desktop config
npx marker-mcp install-config --local

📦 Installation Methods

1. NPX (Recommended)

No installation needed - run directly:

npx @marker-pdf/mcp-server start

2. Global Installation

npm install -g marker-mcp
marker-mcp start

3. Local Installation

npm install marker-mcp
npx marker-mcp start

4. Docker

# Quick start with Docker
docker run -p 3000:3000 marker-mcp

# Or use Docker Compose
git clone https://github.com/VikParuchuri/marker.git
cd marker/node-mcp
docker-compose up

🔧 Configuration

Environment Variables

# Marker server URL (default: http://localhost:8000)
export MARKER_URL=http://your-marker-server:8000

# GPU configuration (optional)
export CUDA_VISIBLE_DEVICES=0
export TORCH_DEVICE=cuda

# LLM configuration for enhanced processing (optional)
export OLLAMA_BASE_URL=http://localhost:11434
export OLLAMA_MODEL=mistral:7b-instruct
export GOOGLE_API_KEY=your_api_key_here

Claude Desktop Integration

Option 1: Auto-generate config

npx marker-mcp install-config --local

Option 2: Manual config

Add to your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%/Claude/claude_desktop_config.json Linux: ~/.config/claude/claude_desktop_config.json

{
  "mcpServers": {
    "marker-pdf": {
      "command": "npx",
      "args": ["marker-mcp", "start", "--local"],
      "env": {
        "CUDA_VISIBLE_DEVICES": "0",
        "TORCH_DEVICE": "cuda"
      }
    }
  }
}

For remote Marker server:

{
  "mcpServers": {
    "marker-pdf": {
      "command": "npx", 
      "args": ["marker-mcp", "start"],
      "env": {
        "MARKER_URL": "http://your-marker-server:8000"
      }
    }
  }
}

🛠️ Available Tools

Once configured, Claude can use these tools:

1. convert_document

Convert documents (PDF, DOCX, PPTX, etc.) to markdown, JSON, or HTML.

Usage: "Convert this PDF to markdown" or "Convert this document to JSON format"

2. extract_tables

Extract tables from documents in various formats.

Usage: "Extract all tables from this PDF as markdown"

3. ocr_document

Perform OCR on scanned documents or images.

Usage: "OCR this scanned document" or "Extract text from this image"

4. start_marker_server

Start a local Marker server (if not running remotely).

Usage: "Start the local Marker server on port 8000"

5. check_marker_status

Check if the Marker server is running and healthy.

Usage: "Check if Marker is running"

🐳 Docker Deployment

Quick Start

# Run with basic setup
docker run -p 3000:3000 -p 8000:8000 marker-mcp

# Run with GPU support
docker run --gpus all -p 3000:3000 -p 8000:8000 marker-mcp

Docker Compose (Production)

git clone https://github.com/VikParuchuri/marker.git
cd marker/node-mcp

# Start basic services
docker-compose up

# Start with LLM support (Ollama)
docker-compose --profile with-llm up

# Production setup with nginx
docker-compose --profile production up

Environment Variables for Docker

Create a .env file:

CUDA_VISIBLE_DEVICES=0
TORCH_DEVICE=cuda
OLLAMA_BASE_URL=http://ollama:11434
OLLAMA_MODEL=mistral:7b-instruct
GOOGLE_API_KEY=your_api_key_here

📋 Command Line Usage

# Start MCP server
marker-mcp start

# Start with local Marker server
marker-mcp start --local --port 8000

# Start with remote Marker server  
marker-mcp start --url http://remote-server:8000

# Run in Docker
marker-mcp docker --port 3000

# Generate Claude config
marker-mcp install-config --remote http://your-server:8000

🔍 Usage Examples

After configuring Claude Desktop, you can ask Claude:

• "Convert this PDF to markdown" - Converts document to markdown format
• "Extract tables from this document as JSON" - Extracts structured table data
• "OCR this scanned document" - Extracts text from images/scans
• "Convert this PowerPoint to HTML" - Converts presentations
• "Process only pages 1-5 of this PDF with OCR" - Selective processing

🚀 Remote Deployment Options

1. Cloud Server

Deploy on any cloud provider (AWS, GCP, Azure, etc.):

# On your server
git clone https://github.com/VikParuchuri/marker.git
cd marker/node-mcp
npm install
npm start

2. Docker on Cloud

docker run -d --name marker-mcp -p 80:3000 marker-mcp

3. Kubernetes

apiVersion: apps/v1
kind: Deployment
metadata:
  name: marker-mcp
spec:
  replicas: 1
  selector:
    matchLabels:
      app: marker-mcp
  template:
    metadata:
      labels:
        app: marker-mcp
    spec:
      containers:
      - name: marker-mcp
        image: marker-mcp
        ports:
        - containerPort: 3000

⚡ Performance & GPU Support

For optimal performance:

1. GPU Support: Set TORCH_DEVICE=cuda and CUDA_VISIBLE_DEVICES=0
2. Memory: Ensure sufficient RAM (8GB+ recommended)
3. Storage: Fast SSD storage for model caching
4. Network: Good bandwidth for downloading models on first run

🛠️ Development

# Clone and install
git clone https://github.com/VikParuchuri/marker.git
cd marker/node-mcp
npm install

# Run in development mode
npm run dev

# Build Docker image
npm run docker:build

# Run tests
npm test

📚 API Reference

The MCP server communicates with Claude through the Model Context Protocol. All tools support:

• File paths: Local file system paths
• URLs: Download files from URLs
• Base64 data: Direct base64 encoded file content

🐛 Troubleshooting

Common Issues

1. "Marker server is not running"

   • Use marker-mcp start --local to start local server
   • Or set correct MARKER_URL for remote server

2. "CUDA out of memory"

   • Set TORCH_DEVICE=cpu to use CPU instead
   • Reduce concurrent processing

3. "Permission denied"

   • Check file permissions
   • Run with appropriate user permissions

4. Models not downloading

   • Ensure internet connection
   • Check disk space (models require several GB)

Debug Mode

DEBUG=* marker-mcp start

📄 License

GPL-3.0-or-later - Same as Marker

🤝 Contributing

1. Fork the repository
2. Create your feature branch
3. Commit your changes
4. Push to the branch
5. Create a Pull Request

📞 Support

• Issues: GitHub Issues
• Discord: Marker Discord
• Documentation: Marker Docs