AGENTKIT/WORKS

pack

Agent Guardrail Pack

Nine controls for agents you don't babysit — they refuse the commands that deleted other people's work.

✓ Claude CodeCodex CLIHermes Agent30+ tools via the open standard

✓ = we've run it on that platform · others load the same SKILL.md format but aren't run-tested yet

PRICE$79$49
Buy now — $49

Your coding agent will eventually run something it should not. These nine controls refuse exactly those commands — each derived from one of eighteen documented incidents where an agent destroyed a production database, a repository, or fifteen years of family photos. Installs as a PreToolUse hook that actually refuses, cites the incident it is protecting you from, and stays silent the rest of the time.

$ tree agent-guardrail-pack/✓ actual contents
agent-guardrail-pack/
├── README.md · INSTALL.md · LICENSE · install.json
├── INCIDENTS.md         ← every rule → the documented failure it prevents
├── hooks/
│   ├── guardrail.py     ← the PreToolUse hook (refuses, cites, stays quiet)
│   └── checks/          ← 9 controls, one file each
│       ├── destructive_migration.py · production_target.py
│       ├── rm_path_escape.py · untracked_delete.py
│       └── … 5 more
└── rules/               ← advisory AGENTS.md / CLAUDE.md for non-hook agents

Why it's different

  • ▸ Refuses at the tool call — does not depend on the model choosing to comply
  • ▸ Three-way decisions: denies catastrophes, asks about ambiguity, stays silent otherwise
  • ▸ Every block cites its incident and links to the write-up
  • ▸ Python 3 stdlib only — nothing to install, nothing to trust

Questions, answered

Does this work with Cursor?+

Partly, and the honest answer matters here. Real enforcement needs a PreToolUse hook, which Claude Code has and Cursor does not. Cursor gets rules files — the same nine rules as instructions the agent can ignore. If you are on Cursor you are buying documented rules, not enforcement.

Will it get in my way?+

It is silent on ordinary commands, and every control ships a test proving it permits normal work — `rm -rf node_modules`, `git push`, `prisma db push` against localhost. Ambiguous cases show a permission dialog rather than guessing.

Is this a sandbox?+

No, and do not treat it as one. It reads the command about to run, so it will not see a deletion inside a script the agent wrote and then executed. It raises the floor against the documented failure modes.

Kits that pair with this one

Each workflow kit teaches a different job. They share the same skill standard, so they stack.