Gives you post-mortem tools for multi-agent systems. Agents call `report()` for heartbeats and `trace()` for actions, linking them by trace ID and parent event. When something fails, `correlate()` walks backward through the chain to find the root cause. Exposes 13 MCP tools including `agentwatch_cascade`, `agentwatch_replay`, and `agentwatch_dashboard` so Claude can diagnose why your swarm fell over. Stores everything in local SQLite, works with any agent framework. The CLI lets you run `npx @nicofains1/agentwatch demo` to see a 5-agent cascade failure traced soup to nuts. Useful if you're running CrewAI, AutoGen, or custom agents and need forensics when one failure triggers three others.
claude mcp add --transport stdio nicofains1-agentwatch -- npx -y @nicofains1/agentwatch