MCPA2A - Multi-Agent Communication Platform

What Problem Does It Solve?

Modern AI systems need agents that can:

• Communicate with each other - Different AI agents (using different LLMs like Gemini, GPT-4, etc.) need to collaborate
• Access external tools - Agents need to use APIs, databases, and enterprise applications
• Work together seamlessly - No matter what framework they're built with

MCPA2A solves this by combining two protocols:

• A2A (Agent-to-Agent) - For agent communication
• MCP (Model Context Protocol) - For tool access

How It Works

1. Host Agent receives user requests
2. Discovers available agents from registry (A2A protocol)
3. Loads available tools from MCP servers
4. Routes the task to the right agent or tool
5. Returns the response to the user

Project Structure

MCPA2A/
├── app/cli/client.py              # CLI client
├── agents/
│   ├── host_agent/                # Orchestrator (port 10001)
│   └── website_builder/           # Child agent (port 10000)
├── mcp/servers/
│   ├── stdio_server.py            # MCP server implementation
│   ├── arithmetic_server.py       # HTTP MCP server
│   └── terminal/server.py         # Shell command server
├── core/
│   ├── a2a/                       # A2A utilities
│   ├── mcp/                       # MCP utilities
│   └── common/                    # Shared utilities
└── pyproject.toml

Installation

git clone https://github.com/yourusername/MCPA2A.git
cd MCPA2A
python -m venv .venv
source .venv/bin/activate
pip install -e .

Add your GOOGLE_API_KEY to .env file.

Configuration

Register Agents

Edit core/a2a/agent_registry.json:

[
    "http://localhost:10000"
]

Configure MCP Servers

Edit core/mcp/mcp_config.json:

{
  "mcpServers": {
    "terminal_server": {
      "command": "python",
      "args": ["mcp/servers/terminal/server.py"]
    }
  }
}

Usage

# Terminal 1: Start child agent
python -m agents.website_builder --port 10000

# Terminal 2: Start host agent
python -m agents.host_agent --port 10001

# Terminal 3: Send queries
python app/cli/client.py --agent http://localhost:10001

Dependencies

• a2a-sdk
• fastmcp
• google-adk
• uvicorn
• httpx

License

MIT License