Your Cart
Loading

Claude Code Seatbelt Kit

On Sale
$27.00
$27.00
Added to cart

Don't Let Your AI Touch Another File Until You Read This!


Your AI helper can wreck your work in one second. Seatbelt refuses the worst commands. Risky ones are refused too, or put to you as a question if you run Claude Code with approval prompts on.

When you let an AI write your code, it can delete your files or break your website without asking. This add-on checks every command before it runs, refuses the dangerous ones, and refuses or asks about the risky ones. $27 once. Try the kit for 14 days. If you’re not satisfied, email caveman@zagairot.com within 14 days of purchase for a full refund, no questions asked.



Build with a safety net

  • Protect what matters. Flagged destructive commands, credential exposure, and deployment changes are blocked or sent to you for approval.
  • Pick up where you left off. Each session starts with what changed, what ran, what failed, and what was decided.
  • Make “done” prove itself. Checks for untested edits and frontend calls to missing backend routes help catch unfinished work.
  • Get your app across the finish line. Includes /fix-my-app, /finish-end-to-end, and /audit-real-vs-stubbed.
  • Keep control. Local project memory stays on your machine. Your existing hooks and MCP servers are preserved.


Included in your download

Windows and Mac/Linux installers with a live self-test, the Sentarion MCP server V1, policies, skills, and installation support.

You also get the Complete Policy Pack FREE—four additional policies, normally $19, already bundled.

Plus an optional 14-day Sentarion v2 trial with signed audit manifests, adversarial code review, budgets, background jobs, and hosted memory. The Seatbelt keeps working without the trial.


Install once. Keep building.

Requires Python 3.10+; the installer offers to install it. Built on the Apache-2.0 open-source Sentarion MCP engine. You’re buying the complete kit: packaging, policies, skills, installers, and support. Tested on Claude Code 2.1.308. The documented minimum for the required onFailure setting is 2.1.295; earlier supported versions were not separately tested.


Real protection. Honest limits. The gate matches the text of a command against written rules. It catches what the rules describe and nothing else. It is a seatbelt, not a guarantee, and not a sandbox. Keep backups. Enforced by Claude Code: a refusal stops that command (confirmed in live Claude Code sessions with auto-approve on, including a command issued by a subagent). If the Seatbelt cannot read a command, hits an error, finds a broken policy file or takes more than 15 seconds, it refuses. On Claude Code 2.1.295 or newer the installer also makes a hook that cannot start block the command; on an older Claude Code that protection is not available, the installer warns you, and the doctor command reports it as not ready. Guidance to the AI, not enforcement: the reminders every 25 tool calls and the three skills. Not covered: deletion inside a script the agent writes and then runs (python clean.py, bash x.sh, make clean), rm with a glob and no -r, and any command the rules do not describe. Inline python -c / node -e code is checked only for common delete calls.

Version 0.4.4 (this download). What’s new since 0.4.3: the git clean check now catches -fd and -fdx (dry runs stay allowed); find -delete, find -exec rm, xargs rm, rsync --delete, rm -f -r and inline python/node delete calls are now refused; the gate refuses when it cannot read or finish a check instead of letting the command through; on Claude Code 2.1.295 or newer the installer sets the host to block when the hook fails; and the doctor command now reports what it detected, what it verified and what it cannot verify. Full notes are in CHANGELOG.md inside the download.


14-day money-back guarantee: Try the kit for 14 days. If you’re not satisfied, email caveman@zagairot.com within 14 days of purchase for a full refund, no questions asked.

I’m one person building this. If installation gives you trouble, email me directly and I’ll help.

Less rebuilding. More shipping. Get the seatbelt and enjoy riding that vibe again now!


(Sentarion MCP V2 upgrade available @ Zagairot.com)



FAQ


What is it?

A set of rules that sit between your AI coding helper and your computer. The worst commands are refused outright, like wiping a database or deleting everything. The risky ones are put to you as a question if you run with approval prompts on, and refused if auto-approve is on. It also reminds the AI of your rules every so often, so a long session doesn't drift. (That reminder is guidance to the AI, not enforcement.)


Who is it for?

Anyone who lets an AI write or run code on their computer, especially every day.


Which AI tools does it work with?

Claude Code. Cursor is in beta. Codex, Gemini CLI and Grok are not part of this kit. Only Claude Code was tested for version 0.4.4.


What do I need?

Python 3.10 or newer. The installer looks for it and offers to install it if it's missing.


Does it send my code anywhere?

No. It runs on your computer and the log stays on your machine.


Will it catch everything?

No. It only catches what its rules describe. It's a seatbelt, not a guarantee, so keep backups.


What if I don't like it?

Try the kit for 14 days. If you’re not satisfied, email caveman@zagairot.com within 14 days of purchase for a full refund, no questions asked.

Pair it with: the AI Coding Agent Rules Pack ($12): https://payhip.com/b/Laqr2 · the Vibe Coder Launch Pack ($9): https://payhip.com/b/wS0AX · or try the approach free first with the 12 Rules pack: https://payhip.com/b/rd1cM

You will get the following files:
  • ZIP (83KB)
  • ZIP (6KB)