Claude Code & Cursor Security Starter Kit — Vet Agent Skills Before You Run Them

I almost shipped a Claude Code skill that Snyk flagged as "Critical Risk." I hadn't noticed anything wrong just reading the code.That's the whole reason this exists. AI coding agents — Claude Code, Cursor, and whatever comes next — can read and write files, run shell commands, hit the network, and install packages. Most of us install skills, MCP servers, and extensions from GitHub repos and community libraries with basically no vetting beyond skimming the README.There are security tools for this category already. They're built for enterprises with a security team and a compliance process. If you're a solo dev or a small team, none of that infrastructure exists for you — you're on your own, deciding in about ten seconds whether to click "allow."This kit is the lightweight version of that missing layer:scan_skill.py — a single Python file, zero dependencies. Point it at a skill or MCP server folder and it flags shell execution, credential access, obfuscated/base64-encoded code, suspicious network calls, and prompt-injection markers hidden in tool descriptions — ranked by severity, with a --fail-on flag so you can wire it into a pre-commit hook or CI job.CHECKLIST.md — the judgment calls a script can't make: where something came from, whether its permission footprint actually matches what it claims to do, and what to do when you're genuinely not sure. Five minutes, every time.POLICY_TEMPLATE.md — drop this into your repo as SKILLS.md and you've got an actual repeatable process — including a decision log — instead of "we should probably check this stuff" living only in someone's head.Who this is for: solo developers and small teams (2-10 people) using AI coding agents day-to-day, who want a real answer to "is this safe to run" without adopting an enterprise security platform to get it.Who this is not for: if you already have a dedicated AppSec function and tooling like Checkmarx or Lasso Security in place, you don't need this — you've already solved the problem this kit solves.Fair warning: this is a fast triage layer, not a full audit. It'll catch obvious red flags and it will occasionally flag something harmless — it will not catch a sufficiently determined, sophisticated attacker. Treat a clean scan as "nothing obvious jumped out," not a guarantee. For anything touching real production credentials, use this as a first pass, not a last one.Instant download. Works offline. No account, no subscription, no dashboard to log into.

Found on Devadex — the discovery index for independent software the big search engines bury. More from gumroad.

Report this listing