pi-warden

# pi-warden

**Your coding agent says "done". It didn't run the tests.**<br>
**pi-warden catches it, and the agent fixes it without you.**

[](https://www.npmjs.com/package/pi-warden)
[](https://www.npmjs.com/package/pi-warden)
[](https://github.com/DevMortimer/pi-warden/stargazers)
[](LICENSE)
[](https://github.com/earendil-works/pi)
[](https://typesafe.ai)

</div>



<div align="center">

**3,350 sessions · 817 risky actions stopped before they ran · 71% of fake "done"s turned into real test runs**

<sub>Forty-one days of the maintainer's real use across all their projects, not just this one, 2026-09-16 to 2026-10-26. Method and noise: [field report](eval/reports/2026-10-27-field-usage/).</sub>

</div>

## Install

```bash
pi install npm:pi-warden
```

Then `/warden enable` (paste a [TypeSafe](https://console.typesafe.ai) key) and `/warden init` (writes a starter `pi-warden.md`). That's it. No key? The offline guards still run.

### Claude Code: per-edit rules

```bash
npm install --global pi-warden
pi-warden init claude
```

Launch Claude with `PI_WARDEN_ENABLED=1` and the active backend's key variable (`TYPESAFE_API_KEY` for the default backend). Successful `Write` and `Edit` calls receive advisory feedback from your `pi-warden.md` and the existing rules configuration. Claude support covers per-edit rules; the other guards below run inside Pi.

Use `init claude --project` for this project's `.claude/settings.json`, and `uninstall claude [--project]` to remove the entries installed in that scope. Shared settings, keys, history and rules files stay in place. [Configuration, limits and uninstall behavior →](docs/claude-code.md)

## It reports. It doesn't nag.

Most guardrails stop and ask you. pi-warden tells **the agent** what it got wrong, and the agent corrects itse