Skill Scan (Round 0 pre-screen)

Answers one question fast: "Is this skill safe to install?" It runs a purpose-built scanner over a SKILL.md / skill folder / MCP server and flags the skill-specific threats a normal code read misses: prompt injection hidden in the prose, MCP tool-poisoning, a skill that reaches more than it declares, a skill that reads your credentials, and the classic "reads a secret, posts it to a webhook" payload.

It is a pre-filter. A clean scan does not skip the house rule two-round gate for anything you did not write yourself. A dirty scan stops the install right there.

How to run it (the only command you need)

The scanner runs in a throwaway Docker container. Static analysis only (--no-llm): no API keys, nothing leaves the machine except an anonymous package-name lookup to the public CVE database. The target is mounted read-only.

Scan a local skill folder or a single SKILL.md:

# Point TARGET at the skill's directory (or the SKILL.md's directory).
TARGET="/absolute/path/to/the-skill"
docker run --rm -v "$TARGET:/scan:ro" skillspector scan /scan --no-llm