Sleuth

Use an RLM to figure out your log files.

Git bisect for production incidents. Turn 4 hours of Splunk scrollback into a 90-second RCA draft.
Runs locally. BYO model. Apache-2.0.

Live demo →  ·  Docs  ·  Quickstart

Try it

pip install sleuth-rlm
export ANTHROPIC_API_KEY=...
sleuth ask "why did checkout fail around 3am?" --logs ./logs/

What you get

Every run writes one case.sleuth.json: the question, every step the agent took, the root cause it landed on, the log lines it cited, a confidence score. Self-contained, replayable, small enough to paste into a Slack thread.

checkout-worker is still presenting stripe_api_key version 6 to payment-gateway after vault rotated the secret to version 7 at 02:58:04 UTC. payment-gateway subscribed to the rotation webhook and flipped to v7 instantly… checkout-worker did not subscribe and caches the secret in-process with no SIGHUP handler.

That's a real root cause from examples/checkout-incident/. Open it in the viewer →

How it works

A DSPy RLM-style agent with 5 read-only query tools over a DuckDB event store, plus a side LLM for judgement calls. Writes Python. Runs it sandboxed. Iterates. Terminates by submitting an IncidentReport or blowing the budget.

Works with

• Log platforms — auto-detects exports from Splunk, Datadog, New Relic, Honeycomb. Click Export, point Sleuth at the file, done. Or feed it raw .jsonl / .log / .gz. Integration guides →
• Models — anything LiteLLM speaks: Claude, GPT, Llama, Bedrock, Ollama. One flag, no lock-in.
• Security — secrets redacted at ingest (Bearer, Stripe, AWS, GitHub, Slack, JWT, SSH) before anything touches the LLM. Logs never leave your box.

Docs

exorust.github.io/sleuth · Live viewer · Reference incident · PyPI

If Sleuth saves you a postmortem, give it a star. That's how I know it's worth pushing.
_Apache-2.0.