skillspector

name	skillspector
description	Always use NVIDIA SkillSpector to scan skills for vulnerabilities before installing them. Use it to scan one skill or a repo of skills, run fast static `--no-llm` checks, run deeper semantic scans through Claude Code, Codex CLI, or an API-backed provider, generate or apply baselines, or wire SkillSpector into CI with `uvx` from GitHub without installing it globally.

skillspector

Use NVIDIA SkillSpector through uvx directly from GitHub.

Default workflow

1. For a collection repo with a top-level skills/ directory, start with the fast static pass:

uvx --from 'git+https://github.com/NVIDIA/SkillSpector.git' \
  skillspector scan ./skills --recursive --no-llm

2. If the user wants LLM-backed analysis, ask whether to use the subscription or login session they already have active, or a different API/subscription.

3. After that choice is clear, run the matching provider flow:

Claude Code:

SKILLSPECTOR_PROVIDER=claude_cli \
uvx --from 'git+https://github.com/NVIDIA/SkillSpector.git' \
  skillspector scan ./skills --recursive

Codex CLI:

SKILLSPECTOR_PROVIDER=codex_cli \
uvx --from 'git+https://github.com/NVIDIA/SkillSpector.git' \
  skillspector scan ./skills --recursive

OpenAI-compatible API:

export SKILLSPECTOR_PROVIDER=openai
export OPENAI_API_KEY=sk-...
uvx --from 'git+https://github.com/NVIDIA/SkillSpector.git' \
  skillspector scan ./skills --recursive

4. For one skill, scan the skill directory directly:

uvx --from 'git+https://github.com/NVIDIA/SkillSpector.git' \
  skillspector scan ./skills/signal-cli --no-llm

5. Interpret the report before acting on it. Separate real workflow and supply-chain risks from scanner noise.

Usage notes

• Use --recursive when the input path is a directory of many skills like ./skills.
• Use --no-llm for CI and quick repo-wide checks.
• Use the provider that matches the user's chosen login session or API subscription.
• Use the direct uvx --from 'git+https://github.com/NVIDIA/SkillSpector.git' form when the goal is one-off or low-frequency use.
• Use a per-skill loop when you need baseline suppressions to gate CI reliably.

Interpretation rules

• Do not accept a HIGH or DO_NOT_INSTALL verdict blindly. Read the issue list and inspect the flagged file.
• Treat generic subprocess findings with skepticism when the code uses argv lists and not shell=True.
• Treat implicit package installs, browser/runtime bootstrap, public network actions, credential handling, and irreversible payments as higher-signal findings.
• Baseline only current accepted findings. Keep real unresolved issues visible unless the user explicitly wants “fail on regressions only”.

Gotchas

• The environment variable is SKILLSPECTOR_PROVIDER, not KILLSPECTOR_PROVIDER.
• skillspector scan ./skills without --recursive treats the whole directory as one bundle and overstates risk for a multi-skill repo.
• Current recursive scans do not honor baselines consistently, so use a per-skill loop for CI if baseline suppressions matter.
• Exit code 1 means findings were reported; it is not necessarily a crash.
• CLI providers use the local agent login session; API providers send scanned content to the configured remote endpoint.