Skip to content
guilyxPublic

About

A git hook that won't let you push code you can't explain. Quizzes you on your own diff. Works with Claude Code, Codex, Cursor.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

216 stars

Watchers

0 watching

Forks

Latest commit

 

History

180 Commits

Folders and files

grip

A git hook that won't let you push code you can't explain.

CI PyPI Python Docs License: BSD-3-Clause

grip whiteboard: the change drawn as parts and arrows, with a question about a malicious client pinned to it

Before your code leaves your machine, grip reads the diff and asks you about it: why it is built this way, where it fails, what a malicious caller could do, how you know it works. In the terminal, or on a whiteboard of your change in the browser. A model grades your answers out of 100. Below your team's pass mark, the push waits.

It checks the developer, not the code. Let AI write it; just be able to defend it. Follow one real change from commit to push to see what it asks and how it grades.

"If I ask 'why did you do X instead of Y?', 'what happens if this actor behaves maliciously?' [...] you should be able to answer confidently." Mitchell Hashimoto, on X, about the whiteboard defense (not affiliated with grip)

Try it in 30 seconds

No hook, no account, no API key: the fake provider runs the whole flow offline.

pipx install grip-hook                 # or: uv tool install grip-hook
mkdir try-grip && cd try-grip && git init -q
echo 'def add(a, b): return a + b' > math.py && git add math.py
grip quiz --provider fake              # five questions in your terminal
grip whiteboard --provider fake        # the same, drawn in your browser

The offline grader does not read your answers, it only checks their length: twenty characters, or the word "because", gets full marks. Any long answer passes and a row of "idk" fails, so this shows the flow, not the grading. For questions written from your diff and real grading, use a model you have (--provider claude-code, codex, gemini or anthropic), or read the walkthrough.

Nothing is blocked until you install the hook with grip init.

Features

  • Five questions from your diff, graded against a hidden rubric into a Grip Score out of 100. Never line numbers or trivia.
  • grip whiteboard: your change drawn as parts and arrows in the browser, with the questions pinned to it, including what a malicious actor could do. Set hook_mode = "whiteboard" and every push is defended there.
  • Works with what you have: reuse Claude Code, Codex or Gemini CLI as the model (no API key), or use Anthropic, any OpenAI-compatible API, or Ollama locally.
  • Inside coding agents: /grip in Claude Code, Codex, Gemini CLI, Cursor and others, and a gate that keeps the agent from pushing until you pass.
  • Out of the way when it should be: a passed diff is remembered, a retry reuses its questions so only grading costs a call, grip skip lets one push through, CI and non-interactive runs are skipped.

Installation

pipx install grip-hook        # or: uv tool install grip-hook

Python 3.11 or newer, on Linux, macOS or Windows. This installs the grip command and a git grip alias.

Quick start

grip blocking a push with lazy answers, then letting it through after real ones

cd your-repo
grip init        # writes .grip.toml, installs the pre-push hook
git push         # grip asks five questions first

grip init writes provider = "auto": each machine uses Claude Code, Codex or Gemini CLI when one is installed. Otherwise set ANTHROPIC_API_KEY, or see Providers.

Other commands you will use:

grip quiz                  # quiz your staged changes now, no hook needed
grip quiz --branch         # quiz the whole branch since main: your pull request
grip whiteboard            # defend the change in the browser
grip quiz --provider fake  # try the flow offline (grades by length only)
grip skip                  # let the next push through
grip status                # hooks, last score, configuration

The walkthrough follows one real change from git commit to an accepted push.

Configuration

.grip.toml at the root of the repository, committed so the team shares one bar:

passing_score = 70        # 0-100
difficulty = "normal"     # easy | normal | hard
provider = "claude-code"  # auto | anthropic | claude-code | codex | gemini | openai | ollama

Every option, environment variables and the pre-commit framework hook are in the configuration reference.

Documentation

guilyx.github.io/grip: getting started, a walkthrough, coding agents, the whiteboard, how grip works inside, and the honest numbers on what the score does and does not measure.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for the development setup and SECURITY.md for reporting vulnerabilities and what data grip sends. This project follows a code of conduct.

Citation

If grip is part of a study or a course, please cite it; GitHub's "Cite this repository" button reads CITATION.cff.

License

BSD-3-Clause. Copyright (c) 2026, Erwin L.

Star history

Star history of guilyx/grip

About

A git hook that won't let you push code you can't explain. Quizzes you on your own diff. Works with Claude Code, Codex, Cursor.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

216 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages