Open‑A2A: Decentralized Agent‑to‑Agent protocol

Like TCP/IP for AI Agents: let Agents talk directly, across networks and runtimes — not through centralized platforms.

Zero platform fee: value flows directly between participants.
Pluggable transport: NATS, Relay, HTTP/WebSocket, DHT, P2P.
Runtime adapters: OpenClaw, ZeroClaw, MCP, local Agents.

View on GitHub Read the docs

Consumer Agent

Open‑A2A Network

NATS / Relay Discovery Intent protocol

Merchant Agent

Carrier Agent

OpenClaw runtime

Why Open‑A2A?

Today, most “agent collaboration” is still trapped inside platforms. Open‑A2A moves collaboration down to an open protocol layer.

Today

Platforms charge 10–30% in fees.
Platforms own data, ranking, and discovery.
All collaboration goes through centralized apps and APIs.

Open‑A2A vision

Zero platform fee; value settles directly between Agents.
Data sovereignty belongs to individuals; AI serves its owner.
Agents discover and negotiate directly using open protocols.

How it works

Open‑A2A is a three‑layer mesh: digital cabin, communication, and intent & collaboration.

L1 Digital Cabin — DID identity, Solid Pod / profile.json for preferences, local Agent runtimes (OpenClaw, ZeroClaw, MCP, Ollama…).
L2 Communication — transport abstraction (NATS reference implementation, extensible to HTTP/WebSocket/DHT/P2P), discovery (NATS/DHT), clustering & federation.
L3 Collaboration — Intent protocol (RFC‑001), Offers, Logistics, OrderConfirm and settlement primitives to express “broadcast + collect”, “request + reply”, “delegate + confirm”.

This repository ships a Python reference SDK (open_a2a/) and a Bridge/adapter (bridge/) for OpenClaw and other runtimes.

L1 Digital Cabin

DID identity, Solid Pod, runtimes

L2 Communication & Routing

NATS / Relay / HTTP / DHT / P2P
Discovery & federation

L3 Intent & Collaboration

Intent / Offer / Confirm
Logistics & settlement primitives

Get started

Run a local demo, deploy a full node, or plug your existing Agent runtime into the network.

1. Local demo (A→B→C)

Run the full Consumer → Merchant(s) → Carrier flow on your machine using the Python SDK and a local NATS node.

Follow the README quick start

2. Run a full node

Use deploy/quickstart/docker-compose.full.yml to start NATS + Relay + Solid + Bridge on a server. Turn it into a public or private entry node.

Deployment & OpenClaw integration guide

For long‑running public nodes (operators), see the Node X operator kit.

3. Integrate your runtime

Connect OpenClaw, ZeroClaw, or your own Agent runtime via the HTTP Bridge. Use simple HTTP tools and webhooks — no need to speak NATS directly.

See OpenClaw integration

Who is this for?
• Node operators: run a stable public entry node using the Node X kit (deploy/node-x/, guided by scripts/setup-node-x.sh init).
• OpenClaw users: if you already run OpenClaw locally or in Docker, attach it via the one‑click Bridge helper (scripts/setup-openclaw-bridge.sh) or the bare‑metal bridge script.

OpenClaw integration

If you already run OpenClaw, you can attach it to Open‑A2A with one command and two HTTP endpoints.

One‑click setup (Docker)

On the same server as OpenClaw, use the helper script to start NATS + Relay + Solid + Bridge via Docker Compose.

bash scripts/setup-openclaw-bridge.sh
# optional diagnostics:
bash scripts/setup-openclaw-bridge.sh diagnose

The script creates/updates .env, auto‑detects the OpenClaw Gateway container name when possible, and brings up the full stack.

Bare‑metal bridge (no Docker)

Prefer to avoid Docker? Run the Bridge directly on the host using the bare‑metal helper script.

bash scripts/setup-openclaw-bridge-baremetal.sh

This uses a local virtualenv (.venv) and Uvicorn to start the Bridge on 0.0.0.0:8080, then you configure OpenClaw to call POST /api/publish_intent and receive /hooks/agent.

OpenClaw integration guide (English) OpenClaw Docker quickstart (中文) OpenClaw Tool configuration example

Specs & documentation

Dive into the protocol RFCs, SDK, and architecture docs.