How to security-scan Claude Code and Codex skills with NVIDIA SkillSpector before installing them

SkillSpector is NVIDIA's security scanner for agent skills. Before you install a Claude Code, Codex or MCP skill, it checks 71 patterns, including prompt injection, data exfiltration and supply-chain risk, and gives a 0 to 100 risk score. Version 2.13.1 came out on October 23. It's Apache 2.0-licensed.

1. Install it with uv. It needs Python 3.12 or later.

uv tool install git+https://github.com/NVIDIA/skillspector.git

Update later with uv tool update skillspector.

1. Pass the skill as a GitHub URL, a directory, a zip or a SKILL.md file. With --no-llm it runs static analysis only and keeps file contents local. It still looks up dependency names and versions on OSV.dev to find known vulnerabilities.

skillspector scan https://github.com/user/my-skill --no-llm
skillspector scan ./my-skill/ --no-llm

1. For LLM analysis, pick a provider. The default, NVIDIA's build.nvidia.com (nv_build), needs a key; if you're signed in to Claude Code, you can use the local claude CLI with no API key. With LLM analysis on, the skill's file contents go to that provider.

export SKILLSPECTOR_PROVIDER=claude_cli
skillspector scan ./my-skill/

A score of 0 to 20 is SAFE, 21 to 50 is CAUTION, and 51 or higher is DO NOT INSTALL. The exit code is 0 at 50 or below and 1 above, so CAUTION also exits 0. To fail on any finding, add --fail-on-findings.

To call it from inside Claude Code, register it as an MCP server. The stdio transport still has the initialize hang reported in issue #199.

uv tool install --force 'skillspector[mcp] @ git+https://github.com/NVIDIA/skillspector.git'
claude mcp add skillspector -- skillspector mcp