A plugin for OpenClaw that lets OpenClaw agents running on separate servers discover and talk to each other by implementing Google's Agent-to-Agent (A2A) v0.3.0 protocol. It runs as a bidirectional gateway: each instance publishes an Agent Card describing itself, accepts inbound A2A requests and dispatches them to a local OpenClaw agent, and can send outbound messages to configured peers. It is aimed at operators building multi-agent networks that span machines connected over Tailscale, a LAN, or public IPs.
On the protocol side it supports three transports (JSON-RPC, REST, and gRPC) with automatic fallback, SSE streaming for live task status, and all three A2A part types (text, file references or inline files, and structured JSON data), serializing each into plain text before handing it to the underlying OpenClaw agent. Peers can be added statically or found automatically via DNS-SD/mDNS discovery, and messages can be routed to specific peers or agent IDs using pattern-, tag-, or skills-based rules.
A distinguishing design choice is a set of optional "bio-inspired" mechanisms borrowed from cell-signaling biology, off by default and backward-compatible with standard A2A behavior:
- Hill-equation affinity scoring for multi-dimensional peer routing
- A four-state circuit breaker that degrades gradually before fully blocking a failing peer
- Density-aware (quorum-sensing) discovery polling with hysteresis
- Adaptive transport selection that ranks transports by measured success rate and latency
- Michaelis-Menten soft concurrency limiting that adds progressive delay before a hard queue limit
It also includes bearer-token authentication with multi-token rotation, SSRF protections and file-size/MIME limits for file transfers, a durable on-disk task store, a JSONL audit trail and metrics endpoint, a companion DevTools CLI for health checks, sending, streaming, discovery, tracing and benchmarking, and an installable agent skill that walks an AI agent through the full setup.
