clawvet

mcp
Security Audit
Fail
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 11 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in action.yml
  • process.env — Environment variable access in apps/api/drizzle.config.ts
  • process.env — Environment variable access in apps/api/src/db/index.ts
  • process.env — Environment variable access in apps/api/src/queue/scan-queue.ts
  • process.env — Environment variable access in apps/api/src/queue/scan-worker.ts
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Offline AST security scanner & install gate for AI agent skills, rules, hooks, and MCP configs (OpenClaw, Claude Code, Cursor, Codex, Windsurf). Catches credential theft, RCE, and rug-pull updates.

README.md

ClawVet

CI
npm
npm downloads
Live Playground

Static and install-time security scanner for AI agent skills, rules, hooks, and MCP configs (OpenClaw, Claude Code, Cursor, Codex, Windsurf). ClawVet vets SKILL.md and every bundled script it references for prompt injection, credential theft, remote code execution, typosquatting, and rug-pull updates. It runs fully offline by default, with optional LLM (--semantic) and OSV.dev (--osv) passes when you turn them on.

Try it with zero install: clawvet.vercel.app (100% client-side WebAssembly Tree-sitter AST dataflow, nothing uploaded).

ClawVet demo

Quick start

# Install once, then run the copy you installed
npm install -g clawvet

# Audit all installed skills, rules, hooks & MCP configs on your machine (prints Machine Permissions Summary)
clawvet audit --summary

# One-command setup for OpenClaw installPolicy + Claude Code PostToolUse guard (supports --dry-run)
clawvet init

# Scan a single skill (prints grade, Permissions summary, and findings)
clawvet scan ./my-skill/

# JSON output for CI/CD
clawvet scan ./my-skill/ --format json --fail-on high

Install globally rather than running npx clawvet: a fresh npx fetch executes whatever version the registry serves at that moment, and cache cleanup breaks gate configs. Update deliberately with npm update -g clawvet, and verify signed npm provenance with npm view clawvet dist.attestations.

Install-time gate & update checks

clawvet gate hooks into OpenClaw's security.installPolicy so every staged skill is vetted before installation completes, without relying on an agent to remember to scan.

Run clawvet gate --print-config and paste the printed block into your OpenClaw config. It resolves through npm's symlink to the real dist/index.js and node binary, emits source: "exec", and warns on stderr if any path needs chmod go-w.

Risk score Grade scan gate (installPolicy)
0-25 A / B approve allow
26-75 C / D warn warn (stops install and prompts)
76-100 F block block
  • Permissions & rug-pull update diff (on by default): every scan extracts what the skill's code can do (connected sites, credential stores, secret env vars, persistence locations, and proved high-risk behaviours), printed under Permissions by clawvet scan and included as capabilities in JSON. clawvet gate records each allowed install under ~/.local/state/clawvet/installed (using XDG_STATE_HOME or os.homedir(), so existing --print-config blocks get update checks unchanged). On an update for the same owner and slug, the gate asks (warn, naming each new permission) if it adds a secret read, persistence, a high-risk behaviour, or a risky site (the publisher's own GitHub/GitLab account is exempt). Use --strict to ask about every new site on updates, or --no-update-check to disable comparison.
  • Optional LLM judge (--semantic): sends static findings and coverage gaps to Anthropic or any OpenAI-compatible/local endpoint (Ollama, vLLM) on your own key. Every cleared gap must cite a real quote from the skill, and hard security floors still apply.
  • Baselines (--baseline <file-or-dir>): accept reviewed findings bound to file hashes (clawvet scan <skill> --write-baseline <file> --reason "...").

Full setup details, chmod g-w,o-w troubleshooting, coverage profiles (default vs --strict), limits, and known gaps are in docs/gate-and-coverage.md.

CI & pre-commit

GitHub Action (scans a skill, optionally uploads SARIF to code scanning, and posts a sticky PR Permissions Diff comment with a README trust badge):

permissions:
  contents: read
  pull-requests: write     # only for comment-pr
  security-events: write   # only for upload-sarif

steps:
  - uses: actions/checkout@v4
  - uses: MohibShaikh/[email protected]
    with:
      path: skills/my-skill
      fail-on: high
      comment-pr: true
      upload-sarif: true

Pre-commit hook (.pre-commit-config.yaml):

repos:
  - repo: https://github.com/MohibShaikh/clawvet
    rev: v0.18.0
    hooks:
      - id: clawvet
        # args: [--fail-on, medium]   # default is --fail-on high

For a repository of skills, run clawvet audit --dir skills --format json --fail-on high in CI.

How it works

Seven passes turn a skill and its bundled scripts into permissions, findings, and a 0-100 risk score. Passes 1-6 are static, offline, and free; Pass 7 (--semantic) and --osv are opt-in.

Pass Module What it checks
1 skill-parser, sources Frontmatter, CommonMark code fences per bundled file, and executable/shebang files
2 static-analysis Regex rules for RCE, reverse shells, credential theft, OAST canaries, miners, obfuscation, and exfiltration
3 dataflow Tree-sitter AST taint analysis for Python, JavaScript/TypeScript, and shell (never executed): credential or multi-secret exfil, decoded/remote execution, persistence writes, ransomware loops
4 metadata-validator Undeclared binaries/env vars, missing descriptions, invalid semver (hygiene only, zero score weight)
5 dependency-checker, manifests npx -y, global installs, and pinned dependencies for --osv
6 typosquat-detector Levenshtein distance against popular ClawHub skills
7 semantic Opt-in (--semantic) LLM judge weighing static evidence and coverage gaps

Scoring: findings are weighted by severity and confidence (critical 30, high 15, medium 7, low 3) with geometric repeat decay for identical non-obfuscation matches (converging at 1.33x). A context pass downweights dual-use capabilities when no exfiltration or remote-exec sink exists, promotes credential-read + outbound-send pairs and install-me envelopes, and pins disqualifying indicators (such as known C2 IPs) to grade F.

Benchmarks

All untrusted skill benchmarks run inside the no-network Docker sandbox (benchmarks/capability-diff/sandbox.sh).

SkillTrustBench v1.1 (held-out 556-case leaderboard subset)

Configuration Accuracy Macro F1 Malicious F1 Malicious recall
clawvet gate (static) 0.660 0.511 0.815 0.788
clawvet gate --semantic (Claude Haiku 5.5) 0.833 0.770 0.934 0.918

Update rug-pull & real-world friction (v0.16.2 -> v0.17.0, held-out test half)

Metric v0.16.2 v0.17.0
Code-injection rug-pulls caught (gate alone -> gate + diff) 83.9% 95.2%
Mixed rug-pulls caught (gate + diff) - 95.2% (100% tune)
Benign ClawHub updates stopped (gate + diff) 65.5% 34.1%
Benign updates prompting on permission diff N/A 5.1% (12.8% --strict)
ClawHub unique skills hard-blocked / cleanly allowed (212 skills) 37.3% / 41.0% 17.5% / 65.6% (0 regressions)

Full benchmark tables, tune vs test splits, MalSkillBench trade-offs, historical 0.9-0.15 corpora, and prior work are in docs/benchmarks.md.

Monorepo, development & telemetry

The published npm package clawvet contains only packages/cli and packages/shared, with zero database, Redis, or OAuth dependencies.

Package Published Description
packages/cli Yes (npm install -g clawvet) Stateless CLI scanner, multi-agent audit, clawvet init, and OpenClaw install gate
packages/shared Yes (@clawvet/shared) Scanner engine, AST dataflow, capability extractor, patterns
web-demo Static site (optional) Zero-backend browser playground (@clawvet/shared + WebAssembly Tree-sitter)
apps/api No (optional self-hosted) Fastify REST API + BullMQ worker (Postgres, Redis, GitHub OAuth)
apps/web No (optional self-hosted) Next.js 14 dashboard
npm install
npx turbo test --filter=@clawvet/shared --filter=@clawvet/api
Optional self-hosted API / web dashboard & API endpoints
docker-compose up -d
cd apps/api && npm run db:push && npm run dev
cd apps/web && npm run dev

Copy .env.example to .env (DATABASE_URL, REDIS_URL, ANTHROPIC_API_KEY, GITHUB_CLIENT_ID, GITHUB_CLIENT_SECRET).

POST   /api/v1/scans          # Submit skill content for scanning
GET    /api/v1/scans/:id      # Get scan result
GET    /api/v1/scans          # List scans (paginated)
GET    /api/v1/stats          # Public stats
POST   /api/v1/webhooks       # Register webhook
DELETE /api/v1/webhooks/:id   # Remove webhook
GET    /api/v1/auth/github    # GitHub OAuth flow
Opt-in telemetry details (off unless you opt in)

Control with CLAWVET_TELEMETRY=off or on (stored in ~/.clawvet/config.json). Never sent: raw skill names, file contents, source code, file paths, env vars, API keys, or PII.

Field Example Purpose
deviceId a1b2c3d4-... Random UUID, not tied to identity
scanCount 42 Scans run on this device
ts, os, osVersion, cliVersion, environment linux, 0.17.0, production Runtime context (dev/CI filtered out)
skillHash 9f2a... (SHA-256) Hash of skill name (raw name never sent)
riskScore, riskGrade, findingsCount, cached 15, B, 3, false Aggregate scan outcome

License

MIT

Reviews (0)

No results found