Skip to content
PrismaLens Docs

PrismaLens Documentation

AI root-cause investigation from your terminal. Open source, BYO-agent, BYO-key, local-first.
Quickstart View on GitHub

PrismaLens is an open-source (Apache-2.0) command-line tool that investigates a firing alert the way an on-call engineer would: it reads your repo, queries your read-only telemetry, and hands back an ordered-evidence report — hypotheses ranked most to least plausible, each backed by evidence it actually gathered. It runs entirely on your machine. There’s no PrismaLens account and no subscription.

Bring your own agent and key

Section titled “Bring your own agent and key”

PrismaLens doesn’t ship a model. It rents a coding-agent harness — Claude Code today, with more on the way — to do the read-only legwork, and calls a model you configure to write the report. Both are BYO-key: you point it at whichever provider you already pay for (Anthropic, OpenAI, Groq, a local Ollama, any OpenAI-compatible endpoint), and PrismaLens reads the credential fresh from your environment on every run and stores nothing. See Bring your own agent & key.

Why not just point a coding agent at my cluster myself?

Section titled “Why not just point a coding agent at my cluster myself?”

You can — but you’d get a chat transcript, not an investigation. PrismaLens adds the structure that makes the output trustworthy at 3AM:

  • A supervisor, not a single prompt. It decomposes the alert, fans the work out across investigation branches, and reduces the results into one report — instead of one long, meandering agent session.
  • An enforced report shape, not free text. Every run ends in the same ordered-evidence structure — ranked hypotheses, what was ruled out, what was and wasn’t covered — so you read leads, not a wall of tool output. Evidence is ordered, never scored with a fake confidence number. See Ordered evidence and our Sample Report.
  • A read-only posture with an honest fidelity report. The harness runs under a read-only permission posture — enforced outright on the default claude-code harness — optionally backed by an OS-level sandbox, and every report records what boundary was actually applied, not just what you asked for. See Sandboxing & permissions.
  • Unattended, with a queryable history. Point Alertmanager at pl listen and firing alerts investigate themselves while you sleep — grouped, budget-capped, and delivered to Slack — with every run saved to a local store you can query later.

Get running in a few minutes

Install, run pl doctor, and investigate your first alert. Start with the Quickstart.

Understand the model

How the two-tier engine works, what BYO-agent means, and why the report ranks evidence instead of scoring it. See Concepts.

Reach for a command

The full flag reference for every pl subcommand lives in the CLI reference.

Know what leaves your machine

Local-first and BYO-key — here’s exactly what flows to your model provider, and what doesn’t. See Data & privacy.

PrismaLens is Apache-2.0 licensed — free to use, self-host, and modify.