Skip to content

Tracedown

Find the step that caused your AI agent's failure.

Your tracing tool shows what your agent did. Tracedown tells you which step broke it, shows the evidence, and replays the run with your fix. It is open source (Apache-2.0, every feature free) and works on top of the traces you already have.

A support agent promises a $120 refund; the right answer is $80. Tracedown traces the claim to an outdated refund policy at step 2 and replays the run with the current one.

Try the live demo: the eight demo failures, read-only, nothing to install. The free host sleeps when idle, so the first visit can take a minute.

Start here
Quick start One command: the server, the web app and eight demo failures
Diagnose runs you already have Paste a trace, or import from Langfuse or OpenTelemetry
How Tracedown decides The method, the evidence behind each verdict, and where it goes wrong
Integrations One line for LangGraph, CrewAI, OpenAI Agents, the Vercel AI SDK and nine more
Benchmarks Accuracy on public benchmarks, losses included

What you get for a failed run

  • The root cause: the step where things went wrong, not just where they crashed.
  • The evidence: what Tracedown saw, in words you can check.
  • Who was at fault: the agent, the user, or nobody yet.
  • Each claim checked: every sentence of the answer, with the source it came from, or none.
  • Replay with your fix: re-run from the broken step; earlier steps come from the recording.

GitHub · PyPI · npm · Discussions · Report a wrong diagnosis