Nightshift
An AI on-call engineer that tries to prove itself wrong.
When an alert fires, Nightshift plans an investigation and sends specialist agents through metrics, logs, deploys, feature flags and runbooks in parallel. It tries to disprove its own hypotheses, then hands the on-call human a root cause where every claim links to the evidence behind it.
- 01Every model turn, tool call and approval is a durable step. Kill a worker mid-incident and another resumes from the last finished step — no repeated LLM calls, no restarted deploys.
- 02Fixes are typed actions behind role and blast-radius policy: they wait for a human, execute durably, then check whether the alert metric actually recovered.
- 03With no API key it runs deterministic baseline agents — same tools, same orchestrator, same guardrails, at zero cost.
- 272tests, run against SQLite and Postgres
- 20simulated incidents scored against ground truth
- 6integrations, emulated end to end before anything real
- FastAPI
- Claude API
- Postgres
- MCP
- Next.js
- Playwright
- Docker
- Stripe







