Sleuth
Use an RLM to figure out your log files. Git bisect for production incidents. Point it at your logs, ask a natural-language question, get a signed case file you can share in a Slack channel or link in a PR.
What it is
A Python library and CLI that wraps a DSPy-style recursive-LM agent around your log
files. The agent has 5 read-only tools over a DuckDB event store, can call a side LLM
for judgement, and terminates by writing an IncidentReport with evidence citations
and a confidence score.
What it isn't
- A hosted SaaS. Runs entirely locally. Your logs never leave your box.
- A log aggregator. Bring your own logs (
.log,.jsonl,.ndjson,.gz). - Tied to any model vendor. BYO model via LiteLLM (OpenAI, Anthropic, Bedrock, Ollama).
The case file
Every run produces one .sleuth.json file: question + trajectory + report + evidence +
ground truth (optional). The file is the product. It's deterministic to replay
(sleuth replay case.sleuth.json), renders in the browser viewer,
and is small enough to paste into a GitHub issue.
Works with your log platform
Auto-detects exports from Splunk, Datadog, New Relic, and Honeycomb. Click Export in your platform of choice, hand Sleuth the file, ask a question. No connectors, no API keys, no platform credentials.
Next
- Quickstart — 90-second tour
- Case File Protocol — the v0.1 schema
- Tools — what the agent can call
- BYO Model — provider configuration
- Eval Harness — how to grade a run against ground truth