Delegation MCP

--- title: Delegation MCP emoji: 🚀 colorFrom: blue colorTo: purple sdk: docker pinned: false license: mit short_description: Intelligent Multi-Agent Routing & Guidance tags: - mcp-server - building-mcp-track-enterprise - multi-agent - agent-orchestration --- # 🚀 Delegation MCP Server **Intelligent Multi-Agent Routing & Guidance** []() []() []() []() []() > *Built for the MCP 1st Birthday Hackathon - Winter 2025* ## ⚡ Quick Start ```bash # One command to install and configure everything python install.py ``` **That's it!** Restart Claude Code and start delegating: ``` "scan this codebase for security vulnerabilities" → MCP suggests: "Delegate to Gemini" → Claude executes: gemini scan . "design an authentication architecture" → MCP suggests: "Handle directly (Claude is best)" → Claude executes: (Internal reasoning) "refactor the delegation engine" → MCP suggests: "Delegate to Aider" → Claude executes: aider --message "refactor delegation engine" ``` **Features**: - ✅ **One-command installation** - 30 seconds to full setup - ✅ **Intelligent Routing** - Rules + Capabilities analysis - ✅ **Privacy-First** - Your code never passes through this server - ✅ **Lightweight** - Minimal footprint, no heavy databases - ✅ **Cross-platform** - Windows, Mac, Linux --- ## 🎮 Try the Interactive Demo **[](https://huggingface.co/spaces/Cduplar/multi-agent-mcp)** Experience the routing intelligence in action! Our HF Space demo lets you: ### Interactive Features: - 🧪 **Test Any Query** - See routing decisions in real-time - 📊 **Routing Transparency** - View the complete decision-making process: - Task classification (security, architecture, refactoring, etc.) - Complexity assessment (simple/medium/complex) - Detected keywords and routing reasoning - CLI command that would be executed - ⚙️ **Live Configuration** - Toggle agents and routing strategies to see how settings affect decisions - 💡 **Example Queries** - Simple and complex multi-step scenarios ### Try This: 1. Visit the [HF Space](https://huggingface.co/spaces/Cduplar/multi-agent-mcp) 2. Enter: *"Audit the authentication system for SQL injection, XSS, and CSRF vulnerabilities"* 3. Watch it route to Gemini with full reasoning 4. Disable Gemini in settings → See it route to Claude instead! **Want to test with real agents?** Duplicate the Space and add your API keys! --- ## 🌟 What Is This? A **lightweight MCP server** that acts as a routing intelligence layer for AI coding agents. Instead of executing tasks itself (which creates a bottleneck and security risk), it analyzes your request and **guides** your main agent (like Claude Code) on which tool to use. **Key Insight**: This follows the **Routing Guidance** pattern: 1. **Analyze**: The server analyzes the prompt (e.g., "audit security"). 2. **Route**: It determines the best agent based on your **presets** and **rules**. 3. **Guide**: It returns the *exact command* to run. 4. **Execute**: The client (Claude) executes the command directly. This ensures **zero lock-in**, **maximum privacy**, and **native performance**. --- ## 🎯 The Core Value Proposition ### Problem Developers manually switch between AI agents, losing context and productivity: - Claude for architecture - Gemini for security analysis - Aider for git operations - Copilot for GitHub integration ### Solution **One MCP server that tells your agent who to call:** ``` You → Claude Code → Delegation MCP → "Use Gemini for this" → Claude calls Gemini ``` **You work with ONE agent, but get the power of ALL agents.** --- ## 📦 Installation ### Prerequisites - Python 3.10+ - At least one AI agent CLI installed: - [Gemini CLI](https://github.com/google/generative-ai-cli): `npm install -g @google/gemini-cli` - [Aider](https://aider.chat): `pip install aider-chat` - [Claude Code](https://claude.ai/download): `npm install -g @anthropic-ai/claude-code` - [GitHub Copilot](https://github.com/features/copilot): `npm install -g github/copilot` ### Automated Installation (Recommended) ```bash # Clone repository git clone https://github.com/carlosduplar/multi-agent-mcp.git cd multi-agent-mcp # One-command install python install.py # Or on Unix/Mac bash install.sh ``` The installer will: 1. Check system requirements 2. Discover installed agents 3. Configure Claude Code automatically 4. Verify everything works **Restart Claude Code and you're ready!** --- ## 🎯 How It Works ### Intelligent Routing Guidance We use a hybrid approach to determine the best agent for the job: 1. **Rule-Based Presets**: Your configured rules take priority (e.g., "Always use Gemini for security"). 2. **Capability Analysis**: If no rule matches, we analyze agent capabilities to find the best fit. **Query**: "scan for vulnerabilities" 1. **Check Rules**: Matches `security_audit` preset? -> **Gemini** 2. **Guide**: Return guidance to use Gemini ### Example Interaction **User**: "Audit my authentication code for SQL injection" **Claude Code** calls `get_routing_guidance`: ```json { "query": "Audit auth.py for SQL injection" } ``` **MCP Server** responds: ```json { "decision": "DELEGATE_TO: gemini", "agent": "gemini", "task_type": "security_audit", "cli_command": "gemini \"Audit auth.py for SQL injection\"" } ``` **Claude Code** then executes: ```bash gemini "Audit auth.py for SQL injection" ``` --- ## 🔧 MCP Tools ### `get_routing_guidance` Get routing guidance for a task. Returns which agent should handle it and the exact CLI command to run. ```python { "query": "Audit auth.py for SQL injection" } ``` ### `discover_agents` Automatically discover available CLI agents on the system and register them. ```python { "force_refresh": false # Optional: force re-discovery } ``` ### `list_agents` List all registered agents and their availability status. ### ⚡ Token Overhead One of the key advantages of this MCP server is its **minimal context footprint**. Here's the actual token usage: ``` MCP Tools: ├─ get_routing_guidance: 601 tokens ├─ discover_agents: 584 tokens └─ list_agents: 554 tokens ───────── Total MCP overhead: 1,739 tokens (0.9% of 200k context) ``` **What this means**: - ✅ Less than 1% of your context budget - ✅ Leaves 99%+ for actual code and conversation - ✅ No heavy prompts or bloated instructions - ✅ Intelligent routing without sacrificing context Compare this to running multiple agent instances or complex orchestration frameworks that can consume 10-20% of your context just for coordination overhead. --- ## 🏗️ Architecture ``` ┌─────────────────────────────────────────┐ │ Claude Code (or other MCP client) │ │ - User chats here │ │ - Calls get_routing_guidance │ │ - EXECUTES the returned command │ └──────────────┬──────────────────────────┘ │ MCP Protocol (stdio) ▼ ┌──────────────────────────────────────────┐ │ Delegation MCP Server │ │ - Analyzes task complexity & type │ │ - Checks rules & capabilities │ │ - Returns guidance (NO EXECUTION) │ └──────────────────────────────────────────┘ ``` ### v0.4.0 - Lightweight Architecture **Privacy & Security**: - **No Code Execution**: The server never executes code or commands. It only suggests them. - **No Data Persistence**: No databases or logs of your code are kept by the server. - **Direct Connection**: Your agent talks directly to the delegated tool (e.g., Claude -> Gemini). **Agent Auto-Discovery**: - Automatically detects installed CLI agents (Claude, Gemini, Aider, etc.) - Verifies agent availability - Graceful error handling --- ## 🗂️ Project Structure ``` multi-agent-mcp/ ├── src/delegation_mcp/ │ ├── server.py # MCP server (Routing Guidance) ⭐ │ ├── delegation.py # Routing logic & scoring │ ├── orchestrator.py # Agent registry │ ├── agent_discovery.py # System scanner for agents │ ├── tool_discovery.py # Tool definitions │ ├── config.py # Configuration handling │ ├── cli.py # CLI tools │ └── adapters/ # Agent definitions │ ├── claude.py │ ├── gemini.py │ ├── copilot.py │ └── aider.py ├── tools/ # Tool definitions (JSON) ├── tests/ # Comprehensive tests └── config/ # Default delegation rules ``` --- ## 🚀 Roadmap ### ✅ Phase 1: Foundation (COMPLETE) - MCP server with routing guidance - Capability-based routing - Agent auto-discovery - Production-grade architecture ### 🔜 Phase 2: Intelligence (Q1 2026) - ML-powered routing - Learning from user feedback - Custom agent definitions ### 🔮 Phase 3: Collaboration (Q2 2026) - Complex multi-step workflows - Parallel agent execution guidance --- ## 🤝 Contributing We welcome contributions! Add new agent adapters, improve routing logic, or enhance documentation. --- ## 📄 License MIT License - see [LICENSE](LICENSE) --- ## 🎯 The Vision > **"You work with ONE agent, but get the power of ALL agents."** Today's AI landscape has amazing specialists, but they work in silos. **Delegation MCP changes that.** It's the intelligence layer that lets agents collaborate, creating something greater than the sum of its parts. --- **Built with ❤️ for the MCP ecosystem**