Introduction
paws makes your coding agent load the right skill before it takes an action you care about.
A skill only helps when Claude loads it, and Claude decides that on its own every turn. Often it doesn't. The PR body goes out full of AI tells, the tests get skipped, the bucket has no tags.
paws is a small hook that runs on your machine. You tie a skill to an action, such as a commit, a PR, a Slack message, a file edit or the start of a session. When Claude tries that action without the skill loaded, paws stops the call once and names the skill. Claude loads it and tries again, and the retry goes through.
The deny-once flow
Claude runs git commit -m "Add retry to the upload client"paws denies This action requires the "humanizer" skill. Invoke it with the Skill tool, then retry this exact command.Claude loads humanizerClaude retries git commit -m "Retry failed uploads"paws allows the commit goes throughThe deny costs one round trip. After that, every call the skill covers goes straight through until the skill leaves Claude's context. How the gate works has the details.
What paws guarantees
- The skill was loaded before the action. paws checks the session transcript for it. Whether Claude then follows the skill depends on the skill and the model.
- Your calls are never rewritten. paws doesn't change a tool call's input or read its output. It allows the call, or denies it with a message.
- The gate works offline. The hook that decides each call never uses the network. Nothing leaves your machine until you run
unbound-paws login. - It isn't a sandbox. Rules match the text of a command, not what the shell does with it. paws gets a cooperative agent to load a skill. It won't stop someone who is trying to get around it. The hook runs
unbound-paws gatefrom your PATH, so a repo whose.claude/settings.jsonsetsenv.PATHcould swap it out there. Review that file like you review.paws/.
Only Claude Code is supported right now. paws is free, Apache-2.0, and needs no account.