Peerlane

Peerlane is an AXL-native workspace for verifiable peer-to-peer AI work. A user submits one task from a single-screen UI. The coordinator receives it and dispatches to a researcher agent over AXL. From there, workers hand off directly to each other: research forwards to verify, verify can negotiate clarification with research peer-to-peer, then verify forwards verified claims to analyst. The final report returns to coord with a confidence score. Every message travels through AXL's encrypted Yggdrasil mesh - no centralized message broker sits between the workers. The UI includes a Proof mode showing node pubkeys, A2A Agent Cards, MCP tool intent, message IDs with parent chains, gossip counts, and a copyable verification bundle. Judges can confirm the same route from Docker logs, AXL topology, and the peer registry.

Four separate Node.js processes, each running its own AXL node binary with a unique ed25519 identity. Inter-agent messages are JSON envelopes sent via AXL's HTTP bridge (POST /send, GET /recv) over a Yggdrasil mesh. Peer discovery uses a shared JSON registry file - each agent writes its pubkey on startup and waits for all four roles to register. The coordinator exposes an HTTP API and WebSocket feed for the React frontend (Vite + TypeScript). LLM calls go through the Anthropic SDK (Claude Sonnet 4.6) with a mock fallback for offline demos. Each agent has a role-specific system prompt. The frontend renders a live mesh timeline, result panel with agent contributions, and a proof view with full message trace. Everything runs via Docker Compose: one image builds the AXL Go binary and compiles the TypeScript agents, then four containers share a volume for keys and the registry. Built with: AXL, Node.js, TypeScript, React, Vite, Docker, Anthropic SDK.