ai-coding-gym-mcp 0.2.6

MCP server for AI Coding Gym - fetch and submit coding challenges

AI Coding Gym - MCP Server

Local MCP server for interacting with the AI Coding Gym platform. Provides tools to fetch coding problems and submit solutions.

Features

• /fetch: Fetch a coding problem and clone the repository to your local machine
• /submit: Submit your solution by committing and pushing changes

Quick Start

Installation

Option 1: Install from PyPI

pip install ai-coding-gym-mcp

Option 2: Install from GitHub

pip install git+https://github.com/yourusername/ai-coding-gym-mcp.git

Option 3: Install from source

git clone https://github.com/yourusername/ai-coding-gym-mcp.git
cd ai-coding-gym-mcp
pip install -e .

Configure Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on Mac):

{
  "mcpServers": {
    "ai-coding-gym": {
      "command": "ai-coding-gym-mcp"
    }
  }
}

Configure VS Code (Copilot with MCP)

Add to your VS Code settings (.vscode/settings.json or User Settings):

{
  "github.copilot.chat.codeGeneration.instructions": [
    {
      "text": "Use AI Coding Gym MCP tools for problem solving"
    }
  ],
  "mcp.servers": {
    "ai-coding-gym": {
      "command": "python",
      "args": ["-m", "server"],
      "cwd": "/path/to/ai-coding-gym-mcp"
    }
  }
}

Or if installed via pip, use the executable directly:

{
  "mcp.servers": {
    "ai-coding-gym": {
      "command": "ai-coding-gym-mcp"
    }
  }
}

Usage

Running the MCP Server

The server uses stdio for communication with MCP clients:

python server.py

Or configure it in your MCP client settings (e.g., Claude Desktop).

Tool: /configure

Configure the MCP server with your user ID. This generates an SSH key pair and registers it with the server.

Parameters:

• user_id (required): Your user ID for authentication
• workspace_dir (optional): Default workspace directory (default: "./workspace")

Example:

{
  "user_id": "user_123",
  "workspace_dir": "./workspace"
}

What it does:

1. Generates an SSH key pair locally (stored in ~/.mcp-keys/)
2. Sends the public key to the server
3. Receives your repository name
4. Stores configuration for future use

Tool: /fetch

Fetches a problem from the backend and clones the repository locally. This tool is single-purpose: it does not install dependencies, run tests, or perform any other follow-up actions after fetching.

Parameters:

• problem_id (required): Problem identifier (e.g., "django__django-10097")
• user_id (optional): Your user ID (uses configured value if not provided)
• workspace_dir (optional): Local workspace directory (default: "./workspace")

Example:

{
  "problem_id": "django__django-10097"
}

What it does:

1. Uses your SSH key from /configure to access the repository
2. Clones only the specific problem branch (shallow clone)
3. Sets up the workspace at workspace/{problem_id}/
4. Does not run installs, tests, or any other post-fetch steps

Tool: /submit

Submits your solution by committing changes and pushing to the remote repository.

Parameters:

• problem_id (required): Problem identifier
• user_id (optional): Your user ID (uses configured value if not provided)
• commit_message (optional): Custom commit message
• force_push (optional): Force push the branch (use with caution; default: false)

Example:

{
  "problem_id": "django__django-10097",
  "commit_message": "Fixed the authentication bug",
  "force_push": false
}

What it does:

1. Stages all changes in the working directory (git add -A)
2. Commits with the provided or auto-generated message
3. Pushes to the remote branch using deployment key (optionally force-pushes)
4. Notifies the backend server about the submission

Tool: /reset

Resets a problem repository back to the original setup commit.

Parameters:

• problem_id (required): Problem identifier
• user_id (optional): Your user ID (uses configured value if not provided)

Example:

{
  "problem_id": "django__django-10097"
}

What it does:

1. Finds the original setup commit with a message that starts with Setup SWE-bench instance:
2. Resets the repo to that commit (git reset --hard)
3. Removes untracked files (git clean -fd)

Backend API Endpoints

The MCP server connects to the hardcoded AI Coding Gym server and uses the following endpoints:

POST /api/configure

Request:

{
  "user_id": "user_123",
  "public_key": "ssh-rsa AAAAB3..."
}

Response:

{
  "repo_name": "user_123-swebench"
}

POST /api/submit

Request:

{
  "problem_id": "django__django-10097",
  "user_id": "user_123",
  "commit_hash": "abc123def456...",
  "branch": "django__django-10097-user_123",
  "commit_message": "Fixed the authentication bug",
  "timestamp": "2026-02-03T10:30:00"
}

Response:

{
  "status": "success",
  "message": "Submission received"
}

Security

• User SSH keys are stored in ~/.mcp-keys/ with 600 permissions
• Keys are generated locally and public key is shared with the server
• SSH host key checking is disabled for convenience (consider enabling in production)
• Configuration is cached in memory during the MCP server session

Troubleshooting

"No credentials found for problem_id"

• Run /configure first to set up your credentials
• Then run /fetch to download the problem

"Git clone/push failed"

• Check network connectivity
• Verify deployment key is valid
• Ensure SSH agent isn't interfering

"Directory already exists"

• Remove the existing directory or use a different workspace location

Development

The server uses:

• mcp: Model Context Protocol SDK
• requests: HTTP client for backend API calls
• subprocess: Git command execution with SSH key management

Local Development

# Clone the repository
git clone https://github.com/yourusername/ai-coding-gym-mcp.git
cd ai-coding-gym-mcp

# Install in development mode
pip install -e .

# Run tests (if available)
pytest

# Test the server locally
python server.py

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Submit a pull request

License

MIT