🤖 Collaborative MCP Proxy Server

Multi-AI collaborative analysis system for Claude Desktop and Claude Code using existing login-based MCP servers.

✨ Features

• Multi-AI Collaboration: Integrates Ollama, Gemini CLI, Codex CLI, and Serena MCP
• ARM64 Mac Optimized: Native Apple Silicon performance
• Login-Based Authentication: Uses existing CLI configurations (no API keys needed)
• Privacy-Focused: Local processing with Ollama for sensitive data
• Pressure Vessel Analysis: Specialized engineering analysis capabilities
• Claude Integration: Works with both Claude Desktop and Claude Code

Installation

Prerequisites

• Node.js 18+
• Existing Gemini CLI MCP and Codex CLI MCP installed and logged in
• Claude Desktop or Claude Code

Setup

1. Clone/Create the project:

mkdir collaborative-mcp-proxy
cd collaborative-mcp-proxy
# Copy the files: package.json, index.js, proxy-handler.js

2. Install dependencies:

npm install

3. Make executable:

chmod +x index.js

Configuration

Claude Desktop Configuration

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "collaborative-proxy": {
      "command": "node",
      "args": ["/path/to/collaborative-mcp-proxy/index.js"]
    }
  }
}

Claude Code Configuration

Add to your MCP configuration:

{
  "collaborative-proxy": {
    "command": "node",
    "args": ["/path/to/collaborative-mcp-proxy/index.js"]
  }
}

Usage

Once configured, you can use the collaborative analysis in Claude:

Basic Analysis

Use the collaborate tool to analyze this pressure vessel specification...

Planning Mode

{
  "tool": "collaborate",
  "arguments": {
    "task": "Create analysis plan for pressure vessel design",
    "mode": "plan"
  }
}

Full Analysis Mode

{
  "tool": "collaborate", 
  "arguments": {
    "task": "Analyze pressure vessel compliance with ASME standards",
    "content": "Vessel specifications...",
    "mode": "apply"
  }
}

Review Mode

{
  "tool": "collaborate",
  "arguments": {
    "task": "Review completed analysis",
    "content": "Previous analysis results...", 
    "mode": "review"
  }
}

Collaboration Modes

1. Plan Mode (mode: "plan")

• Creates detailed analysis plan using Gemini
• Identifies objectives, focus areas, and deliverables
• Best for complex tasks requiring upfront planning

2. Apply Mode (mode: "apply") - Default

• Performs full collaborative analysis
• Gemini: Comprehensive analysis and risk assessment
• Codex: Technical implementation and compliance analysis
• Generates synthesized consensus
• Most comprehensive option

3. Review Mode (mode: "review")

• Reviews and validates existing analysis
• Provides quality assessment and improvements
• Best for validation of completed work

How It Works

Architecture

Claude Desktop/Code
       ↓
Collaborative MCP Proxy
       ↓
┌─────────────┬─────────────┐
│ Gemini CLI  │ Codex CLI   │
│ MCP         │ MCP         │
│ (logged in) │ (logged in) │
└─────────────┴─────────────┘

Workflow

1. Request: Claude sends collaboration request to proxy
2. Distribution: Proxy calls individual MCPs via subprocess
3. Collection: Proxy gathers results from each MCP
4. Synthesis: Proxy generates consensus using Gemini
5. Response: Combined analysis returned to Claude

Agent Specializations

• Gemini: System-level analysis, risk assessment, comprehensive evaluation
• Codex: Technical implementation, code quality, standards compliance
• Consensus: Synthesis of all perspectives with unified recommendations

Implementation Details

Subprocess Calling

The proxy server calls existing MCPs as subprocesses, preserving their login sessions:

const geminiProcess = spawn('gemini-cli-command', args);
const codexProcess = spawn('codex-cli-command', args);

Error Handling

• Timeout protection (2 minutes per MCP call)
• Graceful degradation if one MCP fails
• Detailed error logging for debugging

Mock Implementation

Current implementation includes mock responses for demonstration. To connect to real MCPs:

1. Update callGeminiMCP() to spawn actual Gemini CLI process
2. Update callCodexMCP() to spawn actual Codex CLI process
3. Ensure proper JSON-RPC message formatting

Development

Testing

# Start the server in development mode
npm run dev

# Test with manual JSON-RPC calls
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | node index.js

Debugging

The server logs to stderr, so you can monitor activity:

node index.js 2> debug.log

Extending

To add new MCPs or capabilities:

1. Add new methods to ProxyHandler
2. Update tool schema in handleToolsList()
3. Implement subprocess calling logic

Troubleshooting

Common Issues

1. MCP Not Recognized

• Verify claude_desktop_config.json path is correct
• Restart Claude Desktop after configuration changes
• Check file permissions on index.js

2. Subprocess Errors

• Ensure Gemini CLI and Codex CLI are installed and logged in
• Verify MCP command paths are correct
• Check Node.js version (18+ required)

3. Timeout Issues

• Increase timeout in proxy-handler.js if needed
• Check network connectivity for external MCP calls
• Monitor stderr logs for detailed error information

Logging

All server activity is logged to stderr:

# View logs while running
node index.js 2>&1 | grep "MCP Proxy"

License

MIT License - See LICENSE file for details

Contributing

1. Fork the repository
2. Create feature branch
3. Add tests for new functionality
4. Submit pull request

Roadmap

• [ ] Real MCP subprocess integration
• [ ] Configuration file support
• [ ] Advanced workflow orchestration
• [ ] Result caching and persistence
• [ ] Web UI for collaboration management
• [ ] Integration with more AI models