Deploy Now

SaaS Alternative

LangSmith Langfuse Arize AX Galileo

Stars

18

Forks

10

Watchers

0

Tracedown

Tracedown finds the exact step that made your AI agent fail, shows the evidence, and replays the run with a fix to prove the problem is gone. Import failed runs from Langfuse, Arize Phoenix or any OpenTelemetry exporter with no code changes, or add one line to any of 13 agent frameworks, including LangGraph, CrewAI, and OpenAI Agents. A failed run turns into an AI stack trace you can paste into an issue, naming the root cause, the evidence, and whether the agent or the user was at fault. Replay re-runs the agent from the broken step with your change while earlier steps come from the recording, and the fix check repeats each suggestion three times without firing real tools. It also checks each claim against its sources, compares runs across prompt and model versions, groups failures into incidents with Slack, PagerDuty, or GitHub alerts, and reports cost per run, step, and model. Diagnosis is rules first, so it runs instantly and privately without an API key, with an optional LLM judge for sharper accuracy. Ask Claude Code or Cursor why a run failed through the built-in MCP server, and export spans over OTLP to any collector. Start as one container for a team, then scale out with Compose or a Helm chart; SSO, roles, and an audit log come included. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Tracedown
Tracedown
Tracedown
Tracedown
Tracedown
Tracedown
Tracedown

Benefits

  • Pinpoints the step that broke it
  • Walks back from the visible symptom to the step that caused it, such as a model writing 'Friday' where a date belongs, and shows the evidence plus who was at fault.
  • Proves a fix before you ship
  • Replays the failed run three times with a suggested prompt change or a dropped document, then checks that the problem is gone and nothing new broke, without firing real tools.
  • Works with traces you already have
  • Import failed runs from Langfuse, Arize Phoenix, or any OpenTelemetry exporter with no code changes, or add a one-line SDK call to any of 13 supported agent frameworks.
  • Free, private, and fully self-hosted
  • Rules-first diagnosis needs no LLM or API key, and every feature, including SSO, roles, an audit log, and scale-out storage, ships under Apache-2.0 for your own servers.

Features

  • AI stack traces
  • Each failed run gets a pasteable stack trace built from model calls, tool calls, retrievals, memory, and hand-offs, naming the root cause, the evidence, and a fault verdict.
  • Replay with fixes
  • Earlier steps come from the recording while the agent runs live from the broken step; tdn.try_fixes repeats every suggested fix three times and reports which ones work.
  • Trace importers
  • tracedown-import reads Langfuse, Arize Phoenix, OpenInference, OpenLLMetry, and any OTLP exporter, so failed runs from the last day can be diagnosed from the command line.
  • Incidents and drift
  • Failures group into incidents by root cause with Slack, PagerDuty, or GitHub alerts, and drift reports compare runs across prompt and model versions.
  • Claim checking and cost
  • Each claim in an answer is checked against its sources, while cost is shown per run, step, model, and agent with Prometheus metrics served at /metrics.
  • MCP server for assistants
  • tracedown mcp lets Claude Code, Cursor, or any MCP client ask why a run failed and which fix works, using the same diagnosis as the web app.

Apps Similar to Tracedown