MCP Agent API

A Model Context Protocol (MCP) server that bridges MCP clients with AI agents through the Agent API by Coder.

┌─────────────────┐     ┌──────────────┐     ┌─────────────┐
│ MCP Client      │     │ MCP Server   │     │ Agent API   │
│ (Windsurf,      │◄───►│ (This tool)  │◄───►│ (AI Agent)  │
│  Augment, etc.) │     │              │     │             │
└─────────────────┘     └──────────────┘     └─────────────┘

This package enables you to use AI agents like Goose, Aider, and Claude through any MCP-compatible client. It implements the Model Context Protocol specification to provide a standardized interface for controlling multiple AI agents.

🚀 Key Features

• Multi-Agent Support: Control Goose, Aider, Claude, and other agents through a unified interface
• Agent Lifecycle Management: Detect, install, start, stop, and restart agents
• Message Handling: Seamlessly send and receive messages between clients and agents
• Health Monitoring: Monitor agent and API health status
• Flexible Transport: Support for both stdio and SSE transport protocols
• Command-Line Interface: Comprehensive CLI for all operations

📦 Installation

# Using uv (Recommended)
pip install uv
uv pip install mcp-agentapi

# Using pip
pip install mcp-agentapi

# From source
git clone https://github.com/lord-dubious/mcp-agentapi.git
cd mcp-agentapi
./build.sh
uv pip install dist/*.whl

For development installation:

uv pip install -e ".[dev]"

🔌 Client Configuration

Quick Setup

Add to your MCP client configuration (Claude Desktop, Windsurf, Augment, etc.):

{
  "mcpServers": {
    "agent-controller": {
      "command": "python",
      "args": ["-m", "mcp_agentapi", "--transport", "stdio"],
      "env": {
        "TRANSPORT": "stdio"
      }
    }
  }
}

Agent-Specific Configuration

For Goose:

{
  "mcpServers": {
    "goose-agent": {
      "command": "python",
      "args": ["-m", "mcp_agentapi", "--agent", "goose", "--auto-start"],
      "env": {
        "GOOGLE_API_KEY": "YOUR-GOOGLE-API-KEY"
      }
    }
  }
}

For Aider:

{
  "mcpServers": {
    "aider-agent": {
      "command": "python",
      "args": ["-m", "mcp_agentapi", "--agent", "aider", "--auto-start"],
      "env": {
        "OPENAI_API_KEY": "YOUR-OPENAI-API-KEY",
        "AIDER_MODEL": "deepseek"
      }
    }
  }
}

For Claude:

{
  "mcpServers": {
    "claude-agent": {
      "command": "python",
      "args": ["-m", "mcp_agentapi", "--agent", "claude", "--auto-start"],
      "env": {
        "ANTHROPIC_API_KEY": "YOUR-ANTHROPIC-API-KEY"
      }
    }
  }
}

SSE Transport

1. Start the server:

   mcp-agentapi server start --transport sse --port 8080 --agent goose --auto-start

2. Configure your client:

   {
     "mcpServers": {
       "agent-api-server": {
         "transport": "sse",
         "serverUrl": "http://localhost:8080/sse"
       }
     }
   }

🔍 Troubleshooting

Command Not Found

# Verify installation
pip list | grep mcp-agentapi

# Reinstall if needed
pip install mcp-agentapi

Manual Agent Installation

# Goose
curl -fsSL https://github.com/block/goose/releases/download/stable/download_cli.sh | bash

# Aider
pip install aider-chat

# Claude
npm install -g @anthropic-ai/claude-code

API Keys

• Goose: GOOGLE_API_KEY
• Aider: OPENAI_API_KEY, ANTHROPIC_API_KEY, or DEEPSEEK_API_KEY (depending on model)
• Claude: ANTHROPIC_API_KEY

📁 Project Structure

mcp-agentapi/
├── bin/
│   └── mcp-agentapi            # CLI executable
├── mcp_agentapi/               # Main package
│   ├── server.py               # Server implementation
│   └── src/                    # Core modules
│       ├── agent_manager.py    # Agent lifecycle management
│       ├── api_client.py       # Agent API client
│       ├── config.py           # Configuration management
│       └── utils/              # Utility functions

See FILE_STRUCTURE.md for more details.

🔄 Command-Line Interface

mcp-agentapi <command-group> <command> [options]

Server Commands

# Start the server
mcp-agentapi server start --transport stdio --agent goose --auto-start

# Check server status
mcp-agentapi server status

Agent Commands

# List available agents
mcp-agentapi agent list

# Start an agent
mcp-agentapi agent start goose --auto-install

# Switch agents
mcp-agentapi agent switch claude --restart

# Send a message
mcp-agentapi agent send --content "Hello, agent!" --type user

Configuration Commands

# Show configuration
mcp-agentapi config show

# Set configuration
mcp-agentapi config set transport=stdio agent_type=goose

Shortcuts

The CLI provides shortcuts for common commands:

mcp-agentapi list    # Same as agent list
mcp-agentapi start goose    # Same as agent start goose

🤖 Multi-Agent Support

Supported Agents

• Goose: Google's AI agent
• Aider: AI pair programming assistant
• Claude: Anthropic's AI assistant
• Codex: OpenAI's code-focused model
• Custom: Support for custom agents

MCP Tools for Agent Control

When using an MCP client (Claude Desktop, Windsurf, etc.), you can use these tools:

Tool: list_available_agents
Arguments: {}

Tool: switch_agent
Arguments: {"agent_type": "goose"}

Tool: start_agent
Arguments: {"agent_type": "aider", "auto_install": true}

📚 Documentation

For more detailed documentation, see the docs/ directory:

• MCP Server Architecture
• Technical Design
• Integration Guide

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.