clawvet playground

Know what an agent skill can do before it runs.

Paste a SKILL.md, Cursor rule, AGENTS.md or helper script. ClawVet runs its static rule passes plus Tree-sitter taint tracking right here in your browser and tells you which secrets, hosts and startup files the code touches.

1

Source

sample
Runs 100% locally via WebAssembly Tree-sitter. Nothing is uploaded.
Samples
SKILL.md
0 lines · 0 B Ready.
2

Verdict

AST engine idle
?

Pick a sample, paste a skill or fetch a URL to see its grade, permissions and findings.

Audit every skill, rule, hook and MCP server on your machine

The CLI scans whole multi-file bundles. clawvet init adds install-time gating with rug-pull permission diffs, and the GitHub Action posts a Permissions Diff on every PR.

npm i -g clawvet && clawvet audit --summary

Measured, not claimed

Read the launch report →
SkillTrustBench v1.1 · held-out, 556 skills
0.833 acc / 0.770 macro F1
clawvet gate --semantic
  • ClawVet0.833
  • Cisco Skill Scanner0.775
  • NVIDIA SkillSpector0.712
Macro F1: Cisco 0.614, NVIDIA 0.511 (published leaderboard).
Offline rug-pull detection
95.2% caught
Code-injection rug-pulls, at a 5.1% prompt rate on benign updates. No LLM.
Real ClawHub audit · 209 skills
40.2% read undeclared secret env vars
4.8% touch host credential stores