Cheeisepy/preflight-gcp-iam-guardrails
๐ก๏ธ GCP IAM Guardrails for Claude Code โ free sample Stop your AI coding agent from running destructive gcloud, gsutil, and bq commands against your GCP project. This repo is a free sample from the Preflight studio โ catch it before it ships. Letting Claude Code (or any agent) touch a real cloud project is great until the day it runs gcloud projects delete or gsutil rm -r on the wrong bucket. These deny rules make Claude Code refuse those commands outright โ no confirmation, noโฆ See the full description on the dataset page: https://huggingface.co/datasets/Cheeisepy/preflight-gcp-iam-guardrails.
๐ก๏ธ GCP IAM Guardrails for Claude Code โ free sample
Stop your AI coding agent from running destructive `gcloud`, `gsutil`, and `bq` commands against your GCP project. This repo is a free sample from the **Preflight** studio โ catch it before it ships.
Letting Claude Code (or any agent) touch a real cloud project is great until the day it runs gcloud projects delete or gsutil rm -r on the wrong bucket. These deny rules make Claude Code refuse those commands outright โ no confirmation, no "oops."
What's in this free sample
sample-deny-rules.json โ 10 of the highest-impact deny rules from the full pack, as a ready-to-merge settings.json fragment. Drop the permissions.deny array into your project's .claude/settings.json and Claude Code will block:
Install (30 seconds)
- Open
sample-deny-rules.json. - Merge the
permissions.denyarray into your project's.claude/settings.json(create the file if it doesn't exist). - Done. Try asking Claude to delete a bucket โ it refuses.
What the full product adds
The free sample is the "obvious catastrophes" layer. The paid product is the defense-in-depth version:
- 27 deny rules (17 more than here), covering snapshots, disks, networks, firewall rules, logging sinks, service-account keys, and more.
- An `ask` tier for risky-but-recoverable operations (add-iam-policy-binding, stop instance, create SA key) so they pause for human approval instead of being blocked.
- A curated `allow` list so safe read-only commands (
list,describe,get-iam-policy) never prompt โ no guardrail fatigue. - A PreToolUse hook (
block-destructive-gcloud.sh) that catches obfuscated and chained commands the static rules can miss. - A 31-assertion test suite proving every block/allow decision, so you can trust it.
INSTALL.md,SKILL.md, and workedexamples/(allowed vs blocked).
โ Get the full pack on Gumroad
[GCP IAM Guardrails for Claude Code โ $24](https://lindenberg17.gumroad.com/l/hmtcxf?ref=hf)
From Preflight โ guardrails for agentic coding & cloud. See also: [Terraform Plan Reviewer](https://lindenberg17.gumroad.com/l/qodem?ref=hf) ยท [webOS QA Checklist Runner](https://lindenberg17.gumroad.com/l/fhrioj?ref=hf).
