Teach your agents the house rules.
How a commit reads, which packages get in, how a bucket is set up. paws makes your AI coding agent check your rules before it acts.
$ curl -fsSL https:/ / paws.getunbound.ai/ install.sh | bash- agent$ git commit -m "updated stuff, fixed things"
- pawsHold on. Commits follow the house style.Load skill caveman-commit, then retry.
- skillcaveman-commit loaded
- retry$ git commit -m "fix(cart): round totals to cents"imperative subjectConventional Commitsno body needed
- doneCommitted.
Your agents reach for everything.
Git, npm, the cloud, your docs and your tickets each have a right way in your house. paws checks it the moment the agent reaches for one.
Rules that stick.
Your agent reads CLAUDE.md most of the time. paws puts the rule in front of it at the moment it acts, every time.
Writing a commit message
An imperative subject in Conventional Commits style, and a body only when the why isn't obvious.
Adding a package
A check on the package before it goes in, npm ci in existing projects, and install scripts off at first.
Creating an S3 bucket
No public access, encryption on and your org's tags, all before the bucket exists.
How it works
- 1
Write the rule
Each house rule is one Markdown file in .paws/. It says when it applies and which skill to read.
- 2
Run paws once
paws adds itself to your agent's hooks, in a file you can commit. It only runs when the agent does.
- 3
The agent checks before it acts
When an action matches, paws stops it once and names the skill. The agent loads it and tries again.
---
name: aws-tagging
event: pre-tool
matcher:
tool: Bash
command_regex: ^aws s3 mb
action:
type: require-skill
skill: aws-tagging-standards
deny_message: Buckets need tags here. Load aws-tagging-standards, then retry.
---
# AWS tagging
Every bucket carries Owner and Environment tags and
blocks public access. The skill explains how. This
hook makes sure it is read before the bucket exists.Deterministic policy control for AI actions.
Translation
When the agent finds another way, it was your policy that missed it.
Deterministic for whom?
Any security vendor that tells you “we have deterministic policy control for AI actions” is in the business of shifting the blame from their product to your security team.
Agents rarely take the same path twice. A rule that matches today's command misses tomorrow's rewrite, and the miss gets filed under your policy.
paws makes a smaller promise, and keeps it.
- It notices when the agent reaches for something your team has a rule about.
- It hands the agent that rule, so the retry is done your way.
- It shows you which rules loaded, and when.
Start from a preset.
Most houses run on the same rules. Start from one Unbound already ships.
unbound-paws add awsS3 buckets stay private, encrypted and tagged; IAM writes skip wildcards
unbound-paws add secretsStaged changes are scanned for credentials before every commit
unbound-paws add databaseCount the rows and open a transaction before UPDATE or DELETE
unbound-paws add dependenciesLockfile, audit and license check before a package is added
unbound-paws add gitReal command output before every commit and push
unbound-paws add writingPR bodies and commits that read like a person wrote them
Pick skills in the browser
See which skills paws loaded and approve updates. We never get your prompts or files.
Questions people ask
About skills, CLAUDE.md, hooks and what paws actually sees.
How paws fits
What happens when a rule fires
.paws/: when it applies (a tool, a command pattern or a file path) and which skill to load. Commit the folder and everyone on the repo gets the same house rules.Privacy and cost
unbound-paws sync --dry-run prints the exact request.unbound-paws uninstalltakes paws's hook entries out of your settings and removes the skill it installed. Your other settings stay as they were.House rules for one repo are easy. For two hundred, there's Unbound.
paws works on its own, with no account. Unbound ships the same rules to every machine and tells you when one drifts.