Skip to content

Latest commit

 

History

History
198 lines (159 loc) · 7.21 KB

README.md

File metadata and controls

198 lines (159 loc) · 7.21 KB

MCP Connect

License: MIT

███╗   ███╗ ██████╗██████╗      ██████╗ ██████╗ ███╗   ██╗███╗   ██╗███████╗ ██████╗████████╗
████╗ ████║██╔════╝██╔══██╗    ██╔════╝██╔═══██╗████╗  ██║████╗  ██║██╔════╝██╔════╝╚══██╔══╝
██╔████╔██║██║     ██████╔╝    ██║     ██║   ██║██╔██╗ ██║██╔██╗ ██║█████╗  ██║        ██║   
██║╚██╔╝██║██║     ██╔═══╝     ██║     ██║   ██║██║╚██╗██║██║╚██╗██║██╔══╝  ██║        ██║   
██║ ╚═╝ ██║╚██████╗██║         ╚██████╗╚██████╔╝██║ ╚████║██║ ╚████║███████╗╚██████╗   ██║   
╚═╝     ╚═╝ ╚═════╝╚═╝          ╚═════╝ ╚═════╝ ╚═╝  ╚═══╝╚═╝  ╚═══╝╚══════╝ ╚═════╝   ╚═╝

The Model Context Protocol (MCP) introduced by Anthropic is cool. However, most MCP servers are built on Stdio transport, which, while excellent for accessing local resources, limits their use in cloud-based applications.

MCP Connect is a tiny tool that is created to solve this problem:

  • Cloud Integration: Enables cloud-based AI services to interact with local Stdio based MCP servers
  • Protocol Translation: Converts HTTP/HTTPS requests to Stdio communication
  • Security: Provides secure access to local resources while maintaining control
  • Flexibility: Supports various MCP servers without modifying their implementation
  • Easy to use: Just run MCP Connect locally, zero modification to the MCP server
  • Tunnel: Built-in support for Ngrok tunnel

By bridging this gap, we can leverage the full potential of local MCP tools in cloud-based AI applications without compromising on security.

How it works

+-----------------+     HTTPS/SSE      +------------------+      stdio      +------------------+
|                 |                    |                  |                 |                  |
|  Cloud AI tools | <--------------->  |  Node.js Bridge  | <------------>  |    MCP Server    |
|   (Remote)      |       Tunnels      |    (Local)       |                 |     (Local)      |
|                 |                    |                  |                 |                  |
+-----------------+                    +------------------+                 +------------------+

Prerequisites

  • Node.js

Quick Start

  1. Clone the repository 
    git clone https://github.com/EvalsOne/MCP-connect.git
    and enter the directory 
    cd MCP-connect
  2. Copy .env.example to .env and configure the port and auth_token: 
    cp .env.example .env
  3. Install dependencies: 
    npm install
  4. Run MCP Connect 
    # build MCP Connect
npm run build
# run MCP Connect
npm run start
# or, run in dev mode (supports hot reloading by nodemon)
npm run dev

Now MCP connect should be running on http://localhost:3000/bridge.

Note:

  • The bridge is designed to be run on a local machine, so you still need to build a tunnel to the local MCP server that is accessible from the cloud.
  • Ngrok, Cloudflare Zero Trust, and LocalTunnel are recommended for building the tunnel.

Running with Ngrok Tunnel

MCP Connect has built-in support for Ngrok tunnel. To run the bridge with a public URL using Ngrok:

  1. Get your Ngrok auth token from https://dashboard.ngrok.com/authtokens
  2. Add to your .env file: 
    NGROK_AUTH_TOKEN=your_ngrok_auth_token
  3. Run with tunnel: 
    # Production mode with tunnel
npm run start:tunnel

# Development mode with tunnel
npm run dev:tunnel

After MCP Connect is running, you can see the MCP bridge URL in the console.

API Endpoints

After MCP Connect is running, there are two endpoints exposed:

  • GET /health: Health check endpoint
  • POST /bridge: Main bridge endpoint for receiving requests from the cloud

For example, the following is a configuration of the official GitHub MCP:

{
  "command": "npx",
  "args": [
    "-y",
    "@modelcontextprotocol/server-github"
  ],
  "env": {
    "GITHUB_PERSONAL_ACCESS_TOKEN": "<your_github_personal_access_token>"
  }
}

You can send a request to the bridge as the following to list the tools of the MCP server and call a specific tool.

Listing tools:

curl -X POST http://localhost:3000/bridge \
     -d '{
       "method": "tools/list",
       "serverPath": "npx",
       "args": [
         "-y",
         "@modelcontextprotocol/server-github"
       ],
       "params": {},
       "env": {
         "GITHUB_PERSONAL_ACCESS_TOKEN": "<your_github_personal_access_token>"
       }
     }'

Calling a tool:

Using the search_repositories tool to search for repositories related to modelcontextprotocol

curl -X POST http://localhost:3000/bridge \
     -d '{
       "method": "tools/call",
       "serverPath": "npx",
       "args": [
         "-y",
         "@modelcontextprotocol/server-github"
       ],
       "params": {
         "name": "search_repositories",
         "arguments": {
            "query": "modelcontextprotocol"
         },
       },
       "env": {
         "GITHUB_PERSONAL_ACCESS_TOKEN": "<your_github_personal_access_token>"
       }
     }'

Authentication

MCP Connect uses a simple token-based authentication system. The token is stored in the .env file. If the token is set, MCP Connect will use it to authenticate the request.

Sample request with token:

curl -X POST http://localhost:3000/bridge \
     -H "Authorization: Bearer <your_auth_token>" \
     -d '{
       "method": "tools/list",
       "serverPath": "npx",
       "args": [
         "-y",
         "@modelcontextprotocol/server-github"
       ],
       "params": {},
       "env": {
         "GITHUB_PERSONAL_ACCESS_TOKEN": "<your_github_personal_access_token>"
       }
     }'

Configuration

Required environment variables:

  • AUTH_TOKEN: Authentication token for the bridge API (Optional)
  • PORT: HTTP server port (default: 3000, required)
  • LOG_LEVEL: Logging level (default: info, required)
  • NGROK_AUTH_TOKEN: Ngrok auth token (Optional)

Using MCP Connect with ConsoleX AI to access local MCP Server

The following is a demo of using MCP Connect to access a local MCP Server on ConsoleX AI:

demo-mcp.mp4

License

MIT License