GitHub MCP Server - Multi-Repository Support

🧾 Overview

A production-ready GitHub PR management server with MCP (Model Context Protocol) support for coding agents. Features multi-repository support with dedicated ports for better client compatibility and process isolation.

Key Features

• Multi-Repository Support - Manage multiple repositories simultaneously
• Dedicated Ports - Each repository gets its own port for better MCP client compatibility
• Process Isolation - Repository issues don't affect other repositories
• Master-Worker Architecture - Robust process management with automatic restarts
• Clean MCP Endpoints - Simple URLs: http://localhost:8081/mcp/
• Production Ready - Proper logging, error handling, and service management

---

🏗️ Architecture

Multi-Port Architecture (Current)

The system uses a master-worker architecture where each repository runs on its own dedicated port:

• Master Process (mcp_master.py) - Spawns and monitors worker processes
• Worker Processes (mcp_worker.py) - Handle individual repositories
• Clean URLs - No complex routing: each repository has its own endpoint
• Process Isolation - Issues in one repository don't affect others

VSCode/Amp Integration

Each repository automatically gets a .vscode/settings.json file configured with the correct MCP server endpoint:

{
  "amp.mcpServers": {
    "github-mcp-server": {
      "url": "http://localhost:8081/mcp/"
    }
  }
}

This allows Amp (or other MCP clients) to automatically connect to the repository's dedicated GitHub MCP server.

🚀 Quick Start

1. Install as System Service

# Clone repository
git clone <your-repo>
cd github-agent

# Install dependencies and setup development environment
./setup/setup_system.sh

# Install as system service
# Linux (requires sudo)
sudo ./scripts/install-services.sh

# macOS (installs to user space)
./scripts/install-services.sh

2. Configure Service

# Linux
sudo nano /opt/github-agent/.env

# macOS  
nano ~/.local/share/github-agent/.env

# Set your GITHUB_TOKEN
GITHUB_TOKEN=your_github_token_here

3. Start Service

# Linux
sudo systemctl start pr-agent

# macOS
launchctl start com.mstriebeck.github_mcp_server
# (Service auto-starts on login and auto-restarts if it stops)

4. Manage Repositories

# Linux
sudo /opt/github-agent/.venv/bin/python /opt/github-agent/repository_cli.py list

# macOS
~/.local/share/github-agent/.venv/bin/python ~/.local/share/github-agent/repository_cli.py list

# Add more repositories
# Linux: sudo /opt/github-agent/.venv/bin/python /opt/github-agent/repository_cli.py add <name> <path>
# macOS: ~/.local/share/github-agent/.venv/bin/python ~/.local/share/github-agent/repository_cli.py add <name> <path>

5. Check Status

# Linux
sudo systemctl status pr-agent

# macOS
launchctl list | grep github_mcp

6. Access Your Repositories

Each repository gets its own endpoint and is automatically configured in VSCode:

• Repository 1: http://localhost:8081/mcp/ (configured in .vscode/settings.json)
• Repository 2: http://localhost:8082/mcp/ (configured in .vscode/settings.json)
• Health checks: http://localhost:8081/health

---

🤖 Claude Code Integration

Prerequisites

• GitHub repository checked out locally
• Python 3.8+ and dependencies installed via pip install -r requirements.txt
• Claude Code freshly installed from https://claude.ai/code
• GitHub token with repo scope permissions

Step 1: Configure Environment

# Navigate to your github-agent directory
cd /path/to/github-agent

# Create/edit .env file with your GitHub token
echo "GITHUB_TOKEN=your_github_token_here" > .env
# Or edit .env file manually to set your token

Step 2: Configure Repositories

Edit the repositories.json file to configure your repositories:

{
  "repositories": {
    "my-project": {
      "workspace": "/path/to/your/project",
      "port": 8081,
      "description": "My project repository",
      "language": "python",
      "python_path": "/path/to/github-agent/.venv/bin/python",
      "lsp_server": "pylsp"
    }
  }
}

Step 3: Start the MCP Server

# Deploy and start the server using the deployment script
./scripts/deploy.sh

The server will start and show output like:

🚀 GitHub MCP Master starting up...
✅ Worker for 'my-project' started on port 8081
🎯 Master server is running. Workers are active.

Step 4: Configure Claude Code (Workspace-Specific)

Create workspace-specific MCP configuration files for each repository. This ensures Claude Code only connects to the appropriate MCP server when working in each repository.

For each repository, create a .mcp.json file in the repository root:

In your github-agent directory:

cat > /path/to/github-agent/.mcp.json << 'EOF'
{
  "mcpServers": {
    "github-agent": {
      "type": "http",
      "url": "http://localhost:8081/mcp/"
    }
  }
}
EOF

In your other repositories (if configured):

# Example for a second repository on port 8082
cat > /path/to/your-other-repo/.mcp.json << 'EOF'
{
  "mcpServers": {
    "your-other-repo": {
      "type": "http",
      "url": "http://localhost:8082/mcp/"
    }
  }
}
EOF

Step 5: Verify Connection

1. Open Claude Code in your repository directory (e.g., /path/to/github-agent)
2. Claude Code will prompt you to allow the new MCP server - click "Allow"
3. Test the connection by asking Claude Code something like:

Can you check the current Git branch and find any associated pull requests?

Claude Code should now be able to use tools like:

• git_get_current_branch - Get current Git branch
• github_find_pr_for_branch - Find PR for the current branch
• github_get_pr_comments - Get PR comments
• github_post_pr_reply - Reply to PR comments
• github_get_build_status - Get CI build status
• search_symbols - Search for functions, classes, and variables
• find_definition - Find symbol definitions using LSP
• find_references - Find all references to a symbol
• find_hover - Get detailed hover information for symbols
• And more...

Step 6: Multiple Repositories (Optional)

To work with multiple repositories in Claude Code, edit your repositories.json:

{
  "repositories": {
    "frontend": {
      "workspace": "/path/to/frontend",
      "port": 8081,
      "description": "Frontend app",
      "language": "javascript",
      "python_path": "/path/to/github-agent/.venv/bin/python"
    },
    "backend": {
      "workspace": "/path/to/backend",
      "port": 8082,
      "description": "Backend API",
      "language": "python",
      "python_path": "/path/to/github-agent/.venv/bin/python",
      "lsp_server": "pylsp"
    }
  }
}

Then restart the server with ./scripts/deploy.sh and create .mcp.json files in each repository:

In /path/to/frontend/.mcp.json:

{
  "mcpServers": {
    "github-frontend": {
      "type": "http",
      "url": "http://localhost:8081/mcp/"
    }
  }
}

In /path/to/backend/.mcp.json:

{
  "mcpServers": {
    "github-backend": {
      "type": "http", 
      "url": "http://localhost:8082/mcp/"
    }
  }
}

Troubleshooting Claude Code Integration

Connection Issues:

# Check if the MCP server is running
curl http://localhost:8081/health

# Check MCP endpoint
curl http://localhost:8081/mcp/

# View server logs
tail -f ~/.local/share/github-agent/logs/master.log
tail -f ~/.local/share/github-agent/logs/my-repo.log

Claude Code Debugging:

1. Check Claude Code's developer tools (Help > Toggle Developer Tools)
2. Look for MCP connection errors in the console
3. Verify the MCP configuration file syntax is valid JSON
4. Ensure the repository path in your configuration matches where Claude Code is opened

Common Solutions:

• Restart Claude Code after changing MCP configuration
• Ensure GitHub token has proper repo permissions
• Check that the repository path exists and is accessible
• Verify no firewall is blocking localhost connections

---

⚙️ Configuration

Repository Configuration

The system uses a JSON configuration file (repositories.json) to manage multiple repositories:

{
  "repositories": {
    "my-repo": {
      "path": "/path/to/my-repo",
      "port": 8081,
      "description": "My project repository"
    },
    "other-repo": {
      "path": "/path/to/other-repo", 
      "port": 8082,
      "description": "Another project"
    }
  }
}

CLI Management

Use the repository CLI to manage your configuration:

# List repositories and their status
python3 repository_cli.py list

# Add a new repository  
python3 repository_cli.py add <name> <path> --description="Description"

# Remove a repository
python3 repository_cli.py remove <name>

# Assign ports automatically
python3 repository_cli.py assign-ports --start-port=8081

# Check system status
python3 repository_cli.py status

# Validate configuration
python3 repository_cli.py validate

Environment Variables

# Required
export GITHUB_TOKEN=your_github_token_here

# Optional
export GITHUB_AGENT_REPO_CONFIG=/path/to/repositories.json  # Custom config location
export SERVER_HOST=0.0.0.0                                  # Host to bind to
export GITHUB_AGENT_DEV_MODE=true                          # Enable hot reload

Log Files

Logs are stored in ~/.local/share/github-agent/logs/:

• master.log - Master process logs
• <repo-name>.log - Individual worker logs per repository

---

🔧 Available Tools

Each repository provides these MCP tools:

• git_get_current_branch - Get current Git branch name
• git_get_current_commit - Get current commit information
• github_find_pr_for_branch - Find PR associated with a branch
• github_get_pr_comments - Get all comments from a PR
• github_post_pr_reply - Reply to a PR comment
• github_get_build_status - Get CI/CD build status for commits
• github_check_ci_lint_errors_not_local - Extract linting errors from CI logs
• github_check_ci_build_and_test_errors_not_local - Extract build errors, warnings, and test failures

---

🐛 Troubleshooting

Check Service Status

# Linux
sudo systemctl status pr-agent
sudo journalctl -u pr-agent -f

# macOS
launchctl list | grep github_mcp
tail -f ~/.local/share/github-agent/logs/github_mcp_server.log

View Logs

# Linux
sudo journalctl -u pr-agent -f
tail -f /var/log/github_mcp_server.log

# macOS
tail -f ~/.local/share/github-agent/logs/master.log
tail -f ~/.local/share/github-agent/logs/<repo-name>.log

Restart Service

# Linux
sudo systemctl restart pr-agent

# macOS (auto-restarts due to KeepAlive)
launchctl stop com.mstriebeck.github_mcp_server
# Service will automatically restart

Stop Service Permanently

# Linux
sudo systemctl stop pr-agent
sudo systemctl disable pr-agent

# macOS
launchctl unload ~/Library/LaunchAgents/com.mstriebeck.github_mcp_server.plist

Manual Testing

# Test individual repository endpoints
curl http://localhost:8081/health
curl http://localhost:8082/health

# Test MCP endpoint
curl http://localhost:8081/

Common Issues

1. Service won't start: Check logs and GitHub token in .env file
2. Port conflicts: Restart service or reassign ports via repository CLI
3. Repository not found: Check paths via repository CLI
4. Permission issues (Linux): Ensure www-data has access to repository paths

---

📋 Migration from Single-Port

To migrate from the old single-port server:

1. Stop old service:

   # macOS (permanently stop)
   launchctl unload ~/Library/LaunchAgents/com.mstriebeck.github_mcp_server.plist
   
   # Linux  
   sudo systemctl stop pr-agent
   sudo systemctl disable pr-agent

2. Configure repositories:

   python3 repository_cli.py init --example
   python3 repository_cli.py add <repo-name> <repo-path>
   python3 repository_cli.py assign-ports

3. Start multi-port server:

   python3 mcp_master.py

4. Update MCP client configurations to use new dedicated ports instead of URL routing

---

🔐 Important: Agent User Setup

The code must be submitted by a different user (the "agent" user). GitHub's API has limitations that prevent the PR author from replying to their own review comments.

1. Create a separate GitHub user and invite them to your project
2. Generate a classic GitHub token (not fine-grained) with repo scope
3. Checkout the repository as the agent user
4. Use the agent user's token in GITHUB_TOKEN environment variable

Critical: The agent user must create the pull request, not the main user!

---

📚 Documentation

• Multi-Repository Setup - Detailed multi-repository configuration guide
• HTTP Services Setup - Legacy single-port deployment guide

---

📁 File Structure

Multi-Port Architecture (Current)

• mcp_master.py - Master process that spawns and monitors workers
• mcp_worker.py - Worker process for individual repositories
• repository_manager.py - Repository configuration management
• repository_cli.py - Command-line interface for configuration management
• repositories.json - Repository configuration file

Legacy Single-Port

• github_mcp_server.py - Original unified server with URL routing

Configuration & Deployment

• requirements.txt - Python dependencies
• install-services.sh - Legacy service installation script
• systemd/pr-agent.service - Systemd service file
• config/services.env - Configuration template

Documentation

• README.md - This file (multi-port architecture guide)
• MULTI_REPO_README.md - Detailed multi-repository setup
• HTTP_SERVICES_README.md - Legacy HTTP server guide
• improvements/multi-port-architecture.md - Technical architecture specification

---

🎯 Benefits of Multi-Port Architecture

Technical Benefits

• Better Client Compatibility - Each repository appears as separate MCP server
• Process Isolation - Issues in one repository don't affect others
• Cleaner Architecture - Simplified worker processes, clear separation of concerns
• Easier Debugging - Clear process boundaries and dedicated logs

User Experience Benefits

• Reliable Connections - No more timeout issues with MCP handshakes
• Independent Operation - Can restart/debug individual repositories
• Clean URLs - Simpler endpoint management (localhost:8081/mcp/)
• Better Scalability - Easy to add new repositories

---