Agent Observability
Run the observability server, stream eval results from your test suites, and drop the dashboard React components into your own app.
Agent Observability captures what your voice and chat agents actually do — session recordings, per-turn latency, token usage, tool calls, and eval verdicts — and gives you a dashboard to inspect it all. It receives session reports from the agent-transport SDKs, stores them in Postgres, and serves a dashboard UI you can run as-is or rebuild from the published component registry.
Where do you want to start?
I'm deploying the server
Bring up Postgres and the ingest server with Docker Compose, configure auth and S3, wire up alert webhooks, and read sessions back through the dashboard API.
I'm instrumenting an agent
Ship session reports from your Python or Node agent — automatic on agent-transport workers, one helper call on raw LiveKit workers.
I'm running evals from my test suite
Stream LiveKit-agents eval results into the dashboard straight from pytest or Vitest. Each run lands as an eval run with per-case verdicts.
I'm building a dashboard UI
Install the shadcn-style registry into your own React app: mount the provider once, then compose the hooks, pages, components, and charts.
How the pieces fit
The server is the ingest endpoint and the dashboard host. The Python and Node SDKs instrument your agents to send it session reports; the pytest and Vitest plugins stream eval results from your test runs. Everything you see in the dashboard is built from the published UI component registry — mount the provider, then compose the hooks, pages, components, and charts to build exactly the views you need.