🛡️ SkillAudit

The security layer for AI agent skills. Scan, gate, and enforce policy before your agent installs anything.

43 detection rules · 401 patterns · MCP + A2A coverage · Zero dependencies

# Gate check — should my agent install this?
npx skillaudit gate https://example.com/SKILL.md

# Full scan
npx skillaudit https://example.com/SKILL.md

# Scan MCP manifest for schema poisoning
npx skillaudit manifest tools.json

Why SkillAudit?

AI agents install tools, skills, and MCP servers from untrusted sources. Those skills can steal credentials, exfiltrate data, inject prompts, or manipulate other agents — and most of this is invisible to the user.

SkillAudit catches it. One API call before install. That's it.

Quick Start

1. Gate Check (one line)

The infrastructure endpoint. Returns allow/deny.

curl "https://skillaudit.vercel.app/gate?url=https://example.com/SKILL.md"
# → {"allow": true, "decision": "allow", "risk": "clean", ...}

2. Full Scan

curl "https://skillaudit.vercel.app/scan/quick?url=https://example.com/SKILL.md"

3. Bulk Gate (check multiple skills at once)

curl -X POST https://skillaudit.vercel.app/gate/bulk \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://example.com/skill1.md", "https://example.com/skill2.md"]}'
# → {"allow": false, "denied": 1, "blocked": [...]}

4. Policy Enforcement

curl -X POST https://skillaudit.vercel.app/policy/evaluate-inline \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://example.com/SKILL.md",
    "policy": {"maxRisk": "low", "blockedCategories": ["credential_theft"]}
  }'

What It Detects

43 rule categories, 401 patterns:

Smart context suppression: documentation examples and placeholder tokens are automatically suppressed to minimize false positives.

CLI

Zero install, zero config. Requires Node.js 18+.

# Scan a file, URL, or directory
npx skillaudit SKILL.md
npx skillaudit https://github.com/user/repo
npx skillaudit ./my-agent-project/

# Gate check (CI/CD: exit 0 = allow, exit 1 = deny)
npx skillaudit gate https://example.com/SKILL.md
npx skillaudit gate https://example.com/SKILL.md --threshold high

# Scan MCP manifest for schema poisoning
npx skillaudit manifest tools.json

# CI/CD integration
npx skillaudit SKILL.md --fail-on moderate          # Exit 1 if risk >= moderate
npx skillaudit SKILL.md --markdown >> "$GITHUB_STEP_SUMMARY"  # PR summary
npx skillaudit SKILL.md --json | jq .riskLevel      # Machine-readable

# MCP server mode
npx skillaudit --mcp

API Endpoints

Full interactive docs at skillaudit.vercel.app/docs

Gate (Infrastructure)

Scanning

Policy & Intelligence

Results

Rate limit: 30 req/min per IP. Bypass with API key.

MCP Server

Use SkillAudit as a native tool in Claude Desktop, Cursor, or any MCP client:

{
  "mcpServers": {
    "skillaudit": {
      "command": "npx",
      "args": ["skillaudit", "--mcp"]
    }
  }
}

Tools: skillaudit_gate, skillaudit_scan, skillaudit_scan_content, skillaudit_reputation, skillaudit_batch

GitHub Action

name: SkillAudit
on: [pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx skillaudit . --fail-on high --markdown >> "$GITHUB_STEP_SUMMARY"

CI/CD Integration

# GitHub Actions — gate check before deploy
npx skillaudit gate "$SKILL_URL" --threshold moderate || exit 1

# Generate PR comment
npx skillaudit ./skills/ --markdown > scan-results.md

# Policy enforcement in pipeline
curl -sf -X POST https://skillaudit.vercel.app/policy/evaluate-inline \
  -H "Content-Type: application/json" \
  -d "{\"url\": \"$SKILL_URL\", \"policy\": {\"maxRisk\": \"low\"}}" \
  | jq -e '.pass == true'

Risk Levels

Self-Hosted

git clone https://github.com/megamind-0x/skillaudit
cd skillaudit && npm install && npm start
# → http://localhost:3847

Built by Megamind_0x 🧠

Live App · API Docs · Dashboard · npm