Decentralized P2P runtime for the A2A protocol.
Connect AI agents across any network — no public IP, no VPN, no configuration.

AgentAnycast implements the A2A (Agent-to-Agent) protocol over libp2p peer-to-peer connections. It gives AI agents the ability to securely find and talk to each other across any network — laptops behind NAT, corporate firewalls, or across the internet — with automatic encryption, NAT traversal, and skill-based discovery.

pip install agentanycast    # or: npm install agentanycast

A2A requires every agent to be an HTTP server with a public URL. This excludes agents behind NAT, behind firewalls, and in environments where centralized gateways are unacceptable. AgentAnycast provides the missing transport layer — zero-config connectivity, end-to-end encryption (Noise_XX), automatic NAT traversal, and capability-based routing where the network finds the right agent for you.

from agentanycast import Node, AgentCard, Skill

card = AgentCard(
    name="EchoAgent",
    skills=[Skill(id="echo", description="Echo the input")],
)

async with Node(card=card) as node:
    @node.on_task
    async def handle(task):
        text = task.messages[-1].parts[0].text
        await task.complete(artifacts=[{"parts": [{"text": f"Echo: {text}"}]}])

    await node.serve_forever()

Three ways to reach another agent:

await node.send_task(peer_id="12D3KooW...", message=msg)                # by Peer ID
await node.send_task(skill="translate", message=msg)                    # by skill
await node.send_task(url="https://agent.example.com", message=msg)      # via HTTP bridge

A thin SDK (Python or TypeScript) communicates over gRPC with a local Go daemon that handles P2P networking, Noise_XX encryption, and A2A protocol logic. On a LAN, agents discover each other via mDNS with zero configuration. Across networks, a self-hosted relay provides NAT traversal and hosts a skill registry for capability-based routing — the relay cannot read any traffic.

Built-in adapters for CrewAI, LangGraph, Google ADK, and OpenAI Agents SDK. The daemon also runs as an MCP server for Claude Desktop, Cursor, and ChatGPT.

• Getting Started — Install, run your first agent, connect two agents
• Architecture — Sidecar model, security, NAT traversal, MCP, federation
• Python SDK Reference — Complete API documentation
• Deployment Guide — Production relay, HTTP bridge, metrics, security
• Protocol Reference — Wire format, task lifecycle, gRPC services
• Examples — Skill routing, framework adapters, streaming, LLM agents

• GitHub Discussions — questions, ideas, show & tell
• Contributing Guide — how to get involved