How we review agent skills
Every "Static review" tag on this site means one thing: we read the skill's files and code at a fixed commit and wrote down what it can do on your machine. 85 skills and MCP servers have a review like this today. They are listed on security-reviewed skills.
The steps
- Find the real source. The repository and the exact folder of the skill. The author shown is the GitHub owner of that repository.
- Pin a version. We clone the repository and record the commit hash, its date and the license. The review only describes that commit.
- Classify it. Instructions only (SKILL.md and other Markdown), instructions plus scripts, an MCP server, or a CLI tool.
- Read what it can do. Commands it tells the agent to run; bundled scripts; network calls (we search for http, fetch, curl, requests, urllib, subprocess, eval, exec); files it writes; API keys, tokens or OAuth it needs; packages it installs; whether it acts in your accounts.
- Write the findings. Specific findings with file paths, the risks we see, the review date and the commit. No scores and no "safe" labels.
- Tag the card. One to three tags from a fixed list (instructions only, runs scripts, network access, needs API key, writes files, installs packages, acts in your accounts, offline), taken from the written findings.
What a static review does not cover
- The security review does not run the skill, install its dependencies or call its APIs; pages call it a "static review". A few pages also have a separate "Tested in Antigravity" section: there we installed the skill in Antigravity on a real computer and report only what we saw, with screenshots.
- The review covers the recorded commit only. Later changes in the source repository are not reviewed until we update the page.
- Third-party services a skill talks to (for example Slack or a search API) are not reviewed.
- A static review can miss problems. Before installing, open the skill folder and read its scripts yourself.
If we ever install and run a skill, its page will say so separately, with the date, system, commands and result.
Recently reviewed
- Vulnerability Scanner · 2026-10-23
- Legal Drafting (Contract Review) · 2026-10-23
- DigitalOcean MCP · 2026-10-23
- Cloudflare Wrangler · 2026-10-23
- Playwright CLI · 2026-10-23
- SAG Voice (ElevenLabs TTS) · 2026-10-23
- Content Strategy · 2026-10-23
- ComfyUI · 2026-10-23
- Metasploit Framework · 2026-10-21
- Banner Design · 2026-10-21
- Brand Guidelines · 2026-10-21
- Calendar Management · 2026-10-21
Back to Antigravity skills · Claude Code skills