GitPilot MCP

An autonomous MCP-powered agent that automates GitHub contributions end-to-end, from issue analysis and code changes to pull requests, conflict resolution, and review updates.

🚀 Overview

GitPilot MCP is an Autonomous GitHub Contribution Agent built using the Model Context Protocol (MCP). It helps developers and contributors automate the full open-source workflow:

• Forking repositories

• Analyzing issues

• Making code changes

• Running tests

• Creating pull requests

• Syncing with upstream changes

• Handling merge conflicts

• Updating PRs based on maintainer feedback

The agent is designed to be minimal, safe, and extensible, without overengineering.

✨ Key Features

• 🔁 Fork & clone GitHub repositories

• 🌿 Create and manage feature branches

• 🔄 Sync fork with upstream when new PRs are merged

• 🧠 Semantic code understanding using RAG

• ✍️ Generate minimal, clean git diffs

• 🧪 Run tests and retry fixes automatically

• 📦 Commit & push changes

• 🔀 Create pull requests programmatically

• 🛠️ Update existing PRs after review feedback

• ⚠️ Detect merge conflicts before PR updates

🧠 Architecture (High Level)

Issue → Code Context (RAG) → Patch → Tests → Commit → PR
              ↑
      Sync / Review Feedback

The system follows real-world open-source contribution practices.

📂 Project Structure

mcp-github/
│
├── main.py        # MCP server & tools
├── rag.py         # RAG indexing and search
├── config.py      # GitHub token & workspace
├── .rag/          # Vector store (local)
└── README.md

🧩 Installation & Setup

Make sure you have the following installed:

• Python 3.10+

• Git

• Claude Desktop (latest version)

• Ollama (for embeddings)

Step 1: Install uv

uv is used for dependency management and isolated execution.

pip install uv

Step 2: Clone the Repository

git clone https://github.com/<your-username>/GitPilot-MCP.git
cd GitPilot-MCP

Step 3: Initialize the Project (uv)

Initialize the project environment:

uv init

Step 4: Install Dependencies

Add all required dependencies:

uv add fastmcp gitpython pygithub python-dotenv numpy faiss-cpu ollama

Step 5: For MCP tools testing

uv run fastmcp dev main.py

⚙️ MCP Server Configuration

Use the following configuration to run the MCP server locally via uv + fastmcp in Claude-desktop:

paste this configuration in claudee-desktop-config file.

{
  "mcpServers": {
    "GitPilot-MCP": {
      "command": "<PATH_TO_PYTHON>/Scripts/uv.exe",
      "args": [
        "run",
        "--project",
        "<PATH_TO_PROJECT_ROOT>",
        "--directory",
        "<PATH_TO_PROJECT_ROOT>",
        "fastmcp",
        "run",
        "main.py"
      ],
      "transport": "stdio"
    }
  }
}

🔐 Configuration

Create a config.py file:

GITHUB_TOKEN = "your_github_personal_access_token"
WORKSPACE = "./workspace"

⚠️ The token must have permissions to fork repositories and create pull requests.

🛠️ Available MCP Tools

Some important tools exposed by the agent:

• fork_repo

• clone_repo

• create_branch

• sync_with_upstream

• index_repo

• get_repo_context

• apply_patch

• run_tests

• retry_fix_with_tests

• commit_and_push

• create_pull_request

• check_for_conflicts

• update_pr_after_feedback

Each tool is single-responsibility and composable.

🧪 Testing Workflow

1. Apply generated patch

2. Run tests (pytest -q by default)

3. Retry fixes automatically if tests fail

4. Stop immediately once tests pass

🔄 Handling Maintainer Feedback

When a maintainer requests changes:

• Generate a new patch

• Apply it to the same branch

• Commit & push

The existing PR updates automatically. No new PRs are created.

🧩 Design Principles

• ✅ Minimal diffs

• ✅ Short, inline comments only when necessary

• ✅ No hard refresh or destructive git actions

• ✅ No silent failures

• ✅ Extensible for future improvements

🚧 Limitations

• Repository indexing may be limited on very large repos

• Conflict resolution is assisted, not fully automatic

• Designed for local / controlled environments (remote hardening can be added later)

🔮 Future Improvements

• Scoped repository indexing

• PR comment parsing

• CI status polling

• Refresh-token support

• Remote multi-user MCP deployment

📜 License

MIT License (or as per your choice)

🙌 Acknowledgements

Built using:

• FastMCP

• GitPython

• PyGitHub

• Ollama (for embeddings)

• Qdrant / FAISS (vector search)

👨‍💻 Author

Divyanshu Giri