Scan, don't trust.
Agent skills are code that runs with your permissions. AgentScan reads them and reports what they do — before you install. The verdict is always yours.
About
A scanner for skills you can’t read fast enough.
AgentScan is a local, deterministic security scanner for AI agent skills — Claude Code, Codex, OpenCode, Hermes, and Grok Build. It reads a skill directory and reports what the code does: shell commands, network calls, secrets, licenses, dependencies, and obfuscation. It never runs the skill and never sends data.
It is an independent project. Not affiliated with Anthropic, Snyk, or any vendor. The scanner is free and open source. Trust Pack — ninety audited skills — is a one-time purchase that keeps the project sustainable.
01 — Principles
Deterministic by design.
The scanner has no model and no hallucination. The same input gives the same report, every time. Every finding carries file:line evidence you can check.
Local and private.
Scans run on your machine and never upload the skill. The only optional network call is an OSV vulnerability lookup, and it is off by default.
Open source, MIT.
The scanner's source is public on GitHub. You can read it, build it, fork it, and audit it yourself. Nothing is hidden.
02 — The facts
| License | MIT — scanner |
| Pricing | Free scanner, $59 one-time Trust Pack |
| Telemetry | None. No account, no dashboard, no phone-home. |
| Corpus | Scanned 4,000 public skills for the corpus report |
| Runtimes | Claude Code, Codex, OpenCode, Hermes, Grok Build |
03 — Contact
Found a skill that slipped past the scanner? Have a feature request, a license question, or a security concern? Open an issue or write to the public contact address below. Security reports can go straight to the maintainer.
Transparency note: Trust Pack is sold through Polar, and license keys are validated against Polar’s public API. The scanner itself has no dependency on either.
Scan before you install.
It is free, local, and open source. It takes one command.
$ pipx install agentscan-cli