AI incident investigation on your machine
Your coding agent reads the code. You check the evidence.
curl -fsSL https://prismalens.io/install.sh | sh irm https://prismalens.io/install.ps1 | iex npm install -g prismalens brew install prismalens/tap/prismalens scoop bucket add prismalens https://github.com/prismalens/scoop-bucketscoop install prismalens
PreviewPreviewPreview
Then run pl up
Also
What is AI incident investigation?
AI incident investigation hands something that broke to an AI agent that reads the service's code and works out what went wrong. With PrismaLens, the coding agent you already use does that on your machine, in a fresh git worktree. Its report ranks the hypotheses, each with the evidence for and against it.
How a run goes
-
A signal arrives
An Alertmanager alert, or an incident you write.
-
Your agent digs in
It works in a fresh worktree of the last commit.
-
You stay in charge
Watch it work. Approve or deny what it asks.
-
Read ranked findings
Hypotheses in order, with evidence for and against.
Direct, or through PrismaLens?
| Your agent, asked directly | Your agent, through PrismaLens | |
|---|---|---|
| Starts from | A prompt you write | An Alertmanager alert, or an incident you write |
| Works in | Wherever you run it | A fresh git worktree, never your checkout |
| Answer | The agent's reply | Ranked hypotheses, each with evidence |
| Kept | What the agent keeps | The whole run: steps, findings, what was ruled out |
| Model and login | The agent's own | The agent's own; PrismaLens never calls a model |
Questions, answered
Which coding agents does it drive?
OpenCode, Claude Code, Codex, Gemini CLI and deepagents work out of the box. Other agents that speak the Agent Client Protocol can be added.
Does PrismaLens call a model?
No. Your coding agent does, with its own sign-in, and sends what it reads to its own model provider, as it would if you used it directly.
Why is there no confidence score?
A model has no calibrated way to produce one. The report ranks hypotheses instead, each marked confirmed, supported, speculative or refuted, with the evidence for and against it.
What does it cost?
PrismaLens is free and Apache-2.0 licensed, with no account and no subscription. Your agent uses its own plan or API key.