An A2A-style agent registry: agents publish Agent Cards (name, url, skills, capabilities), heartbeat to stay discoverable, and a brief orchestrator matches required skills (inferred by skills-engine) against registered agents — reporting coverage, recommendations, and explicit skill gaps.
Part of the agentic stack: skills-engine (reasoning over skills) → skills-mcp (tool exposure) →
agent-registry (discovery + orchestration).
| Method | Path | Purpose |
|---|---|---|
| GET | /.well-known/agent.json |
this registry's own A2A-style card |
| POST | /agents |
register an Agent Card |
| GET | /agents |
list live agents (stale filtered) |
| POST | /agents/{name}/heartbeat |
liveness ping |
| DELETE | /agents/{name} |
deregister |
| POST | /briefs |
{topic} → inferred skills → recommended agents + gaps |
- Registry is append-only JSONL with op replay (
upsert/deregister) — crash-safe persistence without a DB - Staleness is time-based (
stale_after_seconds, default 15 min); dead agents drop out of discovery automatically - Briefs never fail silently: uncovered skills are returned as
skill_gapsrather than dropped - Engine outage degrades to HTTP 502 with reason — no fabricated recommendations
MIT