mcp-a2a-bridge 0.1.0

Bidirectional bridge between MCP (Model Context Protocol) and A2A (Agent-to-Agent) protocols

mcp-a2a-bridge

Bidirectional bridge between MCP and A2A protocols

Connect any MCP tool server to the A2A agent ecosystem — and vice versa.

---

The Problem

MCP (Model Context Protocol) and A2A (Agent-to-Agent) are the two dominant agent protocols in 2026 — but they don't talk to each other.

• You have MCP tool servers but need them accessible as A2A agents
• You have A2A agents but need them usable from Claude Desktop or other MCP clients
• You want to mix and match tools from both ecosystems

The Solution

mcp-a2a-bridge provides bidirectional bridging with a single config file:

┌─────────────────┐                        ┌─────────────────┐
│   MCP Servers    │ ──── MCP → A2A ────▶  │   A2A Clients   │
│  (tool servers)  │                        │ (Joule, agents) │
└─────────────────┘                        └─────────────────┘

┌─────────────────┐                        ┌─────────────────┐
│   A2A Agents     │ ──── A2A → MCP ────▶  │   MCP Clients   │
│ (smart agents)   │                        │ (Claude Desktop) │
└─────────────────┘                        └─────────────────┘

Quick Start

Install

pip install mcp-a2a-bridge

MCP → A2A (one command)

Turn any MCP server into an A2A agent:

# Expose a local MCP server as an A2A agent on port 8000
mcp-a2a-bridge quick python -m my_mcp_server --port 8000

# Check it works
curl http://localhost:8000/.well-known/agent-card.json
curl http://localhost:8000/health

A2A → MCP (one command)

Turn any A2A agent into MCP tools:

# Expose an A2A agent as MCP tools
mcp-a2a-bridge expose https://my-agent.example.com --port 9000

Then add to your Claude Desktop config:

{
  "mcpServers": {
    "agent-bridge": {
      "url": "http://127.0.0.1:9000/mcp"
    }
  }
}

Config File (advanced)

For multi-server setups, create a bridge.yaml:

# Connect to multiple MCP servers
mcp_servers:
  - name: my-tools
    transport: stdio
    command: python
    args: ["-m", "my_mcp_server"]

  - name: remote-tools
    transport: streamable-http
    url: http://localhost:3000/mcp

# Expose as a single A2A agent
a2a_agent_name: "Multi-Tool Agent"
a2a_port: 8000

# Also bridge A2A agents to MCP
a2a_agents:
  - name: order-analyst
    url: https://my-agent.example.com

mcp_port: 9000

# LLM for routing natural language → tool calls
llm:
  model: gpt-4o-mini

mcp-a2a-bridge serve --config bridge.yaml

How It Works

MCP → A2A Direction

1. Connects to configured MCP server(s) and discovers all tools
2. Auto-generates an A2A AgentCard with skills derived from MCP tools
3. When an A2A request arrives (natural language), uses an LLM to:
   • Parse the query
   • Select the right MCP tool(s)
   • Format arguments
   • Call the tool(s) via MCP
4. Returns results as A2A artifacts with full task lifecycle

A2A → MCP Direction

1. Fetches AgentCard from configured A2A agent(s)
2. Generates MCP tools from agent skills (one tool per skill)
3. When an MCP call_tool arrives, forwards as an A2A message/send
4. Returns the agent's response as MCP TextContent

Python API

For programmatic use:

import asyncio
from mcp_a2a_bridge import MCPtoA2AAdapter, A2AtoMCPAdapter
from mcp_a2a_bridge.common.config import BridgeConfig

# MCP → A2A
config = BridgeConfig.from_yaml("bridge.yaml")
adapter = MCPtoA2AAdapter(config)
app = await adapter.build_app()  # Returns FastAPI app

# A2A → MCP
adapter = A2AtoMCPAdapter(config)
mcp_server = await adapter.build_server()  # Returns FastMCP server

Architecture

src/mcp_a2a_bridge/
├── cli.py                  # Click CLI (serve, quick, expose)
├── common/
│   ├── config.py           # YAML config loader (BridgeConfig)
│   └── transport.py        # MCP transport helpers (stdio/SSE/HTTP)
├── mcp_to_a2a/
│   ├── adapter.py          # MCPtoA2AAdapter (main orchestrator)
│   ├── agent_card.py       # Auto-generate AgentCard from MCP tools
│   ├── executor.py         # A2A AgentExecutor → MCP tool calls
│   └── tool_router.py      # LLM-based natural language → tool routing
└── a2a_to_mcp/
    ├── adapter.py           # A2AtoMCPAdapter (main orchestrator)
    └── tool_factory.py      # Generate MCP tools from A2A skills

Supported Transports

Protocol	Transport	Status
MCP	stdio (local subprocess)	✅
MCP	SSE (HTTP)	✅
MCP	Streamable HTTP	✅
A2A	JSON-RPC	✅
A2A	REST (HTTP+JSON)	✅
A2A	gRPC	🔄 (install mcp-a2a-bridge[grpc])

LLM Support

The MCP → A2A direction uses litellm for routing, so any LLM works:

llm:
  model: gpt-4o-mini          # OpenAI
  # model: claude-sonnet-4-20250514  # Anthropic
  # model: gemini/gemini-2.5-pro    # Google
  # model: ollama/llama3           # Local via Ollama

Set OPENAI_API_KEY, ANTHROPIC_API_KEY, etc. as environment variables, or pass api_key in config.

Contributing

git clone https://github.com/naveenkumarbaskaran/mcp-a2a-bridge.git
cd mcp-a2a-bridge
pip install -e ".[dev]"
pytest

License

MIT — see LICENSE.

Author

Naveen Kumar Baskaran — GitHub · LinkedIn

---

The answer to "MCP vs A2A" is "both."