Bouncer

Bouncer

Gemini stands at the door. Independent quality gate that audits Claude Code's output before it can stop. Score below threshold? Claude keeps working.

Why it gets shared

- Claude cannot wave through unverified work. Another model checks the evidence.
- One install gives you both the automatic Stop hook and the on-demand /bouncer skill.
- The repo is explicit about security, data flow, and how to test the installer in a clean environment.

Live site

https://buildingopen.github.io/bouncer/

What it looks like

Quick audit:

Deep audit:

How it works

1. Claude Code triggers the Stop hook when it's about to return a response
2. The hook extracts context: user messages from transcript, tool call results, git diff, CLAUDE.md, workplan
3. Everything is sent to Gemini 2.5 Flash for independent scoring (1-10)
4. If score = threshold, Claude is allowed to stop
6. On re-audit (stop_hook_active=true), the hook audits again rather than skipping

Security note

This hook sends the following data to the Google Gemini API:
- Claude's assistant response (up to 200k chars)
- User messages from the conversation transcript (last 3, up to 50k chars total)
- Tool call activity and results from the transcript (evidence of work done)
- Project CLAUDE.md and active workplan
- Git diff of staged and unstaged changes (up to 50k chars)

Review Google's data handling policies before use.

Setup

Quick install (recommended)

This installs google-genai into your Python user site, copies hook + skill files, registers the Stop hook in settings.json, and enables bouncer. You just need to set your API key:

Add this to your .bashrc/.zshrc. Get a free key at aistudio.google.com/apikey. If no key is set, the hook fails open (exits 0, does not block).

Clean uninstall

Manual…