a2a-mcp-bridge
Bridges MCP clients (like Claude Desktop) to A2A agents, enabling message sending and agent card retrieval via a stateless, non-persistent server.
README
a2a-mcp-bridge
A minimal MCP server that bridges MCP clients
(Claude Desktop, etc.) to A2A agents — built directly on
a2a-sdk 0.3.26+, so it speaks the A2A
protocol version that current agent servers (e.g. Google ADK's to_a2a) actually
implement.
Stateless by design: no registry file, no local persistence, nothing written to disk. Each call opens a connection, does the round trip, and returns — so it never hits the file-permission traps that registry/cache-based bridges run into on locked-down or sandboxed clients.
Why this exists
The most visible community bridge, PyPI's a2a-mcp-server, does not depend on
a2a-sdk at all — it vendors a hand-copied client from the pre-1.0 draft A2A
protocol. Its hardcoded JSON-RPC methods are tasks/send, tasks/sendSubscribe,
tasks/get, tasks/cancel, tasks/pushNotification/{get,set}.
Current a2a-sdk (0.3.26+) servers use a different method set entirely:
message/send, message/stream, tasks/get, tasks/cancel,
tasks/pushNotificationConfig/{get,set,list,delete}, tasks/resubscribe,
agent/getAuthenticatedExtendedCard.
Point the old bridge at a modern A2A server and every call fails with
-32601 Method not found — the server has genuinely never heard of tasks/send.
This isn't a version skew you can fix by bumping a pin; the dialect changed.
a2a-mcp-bridge sidesteps the problem by using a2a-sdk's own client
(ClientFactory, send_message), so it always speaks whatever protocol the SDK
you install implements.
Install
Option A — uvx, no clone, no install:
uvx --from git+https://github.com/ytugarev/a2a-mcp-bridge a2a-mcp-bridge
Option B — pip, from source:
git clone https://github.com/ytugarev/a2a-mcp-bridge
cd a2a-mcp-bridge
pip install -e .
Option C — single file, zero install: src/a2a_mcp_bridge/server.py carries a
PEP 723 inline metadata header, so you can
download that one file and run it directly with uv — uv resolves and caches
its dependencies on first launch, no pip install and no venv to manage:
uv run --script /absolute/path/to/server.py
Configure
Two environment variables, both optional:
| Variable | Default | Purpose |
|---|---|---|
A2A_AGENT_URL |
http://localhost:8001 |
Default A2A agent endpoint (also overridable per tool call) |
A2A_TIMEOUT_SECONDS |
300 |
HTTP client timeout — raise this for slow/long-running agents |
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"a2a-bridge": {
"command": "a2a-mcp-bridge",
"env": {
"A2A_AGENT_URL": "http://192.168.1.50:8001"
}
}
}
}
Using the single-file uv run --script form instead (Option C above):
{
"mcpServers": {
"a2a-bridge": {
"command": "/absolute/path/to/uv",
"args": ["run", "--script", "/absolute/path/to/server.py"],
"env": {
"A2A_AGENT_URL": "http://192.168.1.50:8001"
}
}
}
}
Both command and every path in args must be absolute. MCP clients launch
server subprocesses with the working directory set to some system default (on
Windows, Claude Desktop uses C:\WINDOWS\System32) and often a stripped PATH,
so a bare "uv" or a relative "server.py" will silently fail to resolve —
this is the single most common setup error with this bridge (and MCP servers in
general), not a bug in the bridge itself. On Windows, also double check your
actual config path: Store-installed Claude Desktop keeps it under
AppData\Local\Packages\<package-id>\LocalCache\Roaming\Claude\, not the
%APPDATA%\Claude path most docs assume.
Tools exposed
get_agent_card(agent_url?)— fetches the target agent's name, description, and skills from its/.well-known/agent.jsoncard.send_a2a_message(message, agent_url?)— sends a message to the agent and returns its final text response. Runs non-streaming so the client gets one clean answer per call, even if the underlying agent takes minutes.
Both tools accept an optional agent_url override, so a single bridge
instance can talk to multiple agents if needed.
Requirements
- Python 3.10+
- An A2A server speaking
a2a-sdk0.3.26+ semantics (e.g. Google ADK'sto_a2a, or anything else built on the same SDK)
Development
pip install -e ".[dev]"
pytest
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.