GLM-OCR MCP Server

A Model Context Protocol (MCP) server that provides local OCR capabilities using the GLM-OCR model via Ollama. This server enables AI assistants (like Claude Desktop or VS Code with Continue) to extract text, tables, and mathematical formulas from images locally.

🌟 Features

• 📝 Text Extraction: Convert images into clean Markdown text.
• 📊 Table Recovery: Preserve complex table structures in Markdown format.
• 🧮 Math Support: Automatically convert mathematical formulas to LaTeX.
• 🖼️ Broad Support: Works with PNG, JPG, and other standard image formats.
• 🔧 Privacy-Focused: Local processing via Ollama (no cloud dependencies).

📋 Prerequisites

1. Node.js (v18 or higher)
2. Ollama installed and running
3. The GLM-OCR model pulled in Ollama:

ollama pull glm-ocr

🚀 Quick Start (via npx)

You can run this server directly without cloning the repository by using npx.

⚙️ Configuration

Claude Desktop

Add the following to your Claude Desktop configuration file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "glm-ocr": {
      "command": "npx",
      "args": ["-y", "@vinhnguyen/glm-ocr-mcp"],
      "env": {}
    }
  }
}

Continue.dev (VS Code / JetBrains)

Add this to your config.json (usually found at ~/.continue/config.json or ~/.continue/config.yaml):

{
  "mcpServers": {
    "glm-ocr": {
      "command": "npx",
      "args": ["-y", "@vinhnguyen/glm-ocr-mcp"],
      "env": {},
      "description": "GLM-OCR document processing via Ollama",
      "timeout": 120
    }
  }
}

🛠️ Usage

Once configured and Ollama is running, your AI assistant will have a new tool called ocr_document. You can use natural language to trigger it:

• "Extract the text from this screenshot: /path/to/image.png"
• "Read the table in /Users/me/Downloads/invoice.jpg and format it as markdown."
• "What is the math formula in this image? /path/to/notes.png"

How It Works

1. The assistant sends a local image path to the server.
2. The server validates the file and converts it to base64.
3. It communicates with your local Ollama instance using the glm-ocr model.
4. The extracted content is returned directly into your chat.

📂 Manual Installation (For Development)

If you want to modify the server or run it from source:

1. Clone the repository:

   git clone https://github.com/vinhnguyen/glm-ocr-mcp.git
   cd glm-ocr-mcp

2. Install dependencies:

   npm install

3. Run locally for testing:

   npm start

❓ Troubleshooting

Error: "Ollama connection refused"

Ensure Ollama is running in your system tray or run ollama serve in a terminal.

Error: "model 'glm-ocr' not found"

Run ollama pull glm-ocr to download the specific OCR model.

Timeout Issues

OCR processing is GPU/CPU intensive. If the process times out, ensure your configuration has a timeout value of at least 120 seconds.

📄 License

ISC

✨ Acknowledgments

• Model Context Protocol
• Ollama
• GLM-OCR Model

⚠️ Note for Publishing

If this is your first time publishing this scoped package to NPM, remember to use the public access flag:

npm publish --access public