secrets-audit-mcp

MCP server that detects leaked credentials in source code. Zero dependencies. Single file.

Detects API keys, OAuth tokens, private keys, webhooks, and crypto wallet secrets across 30+ providers (AWS, GCP, GitHub, Stripe, OpenAI, Anthropic, Slack, Discord, Telegram, Twilio, SendGrid, Heroku, DigitalOcean, npm, HuggingFace, Replicate, Cloudflare, and more).

Companion to skill-audit-mcp (behavioral patterns) — together they cover secrets + behaviors in one MCP toolchain.

Why

Most secret scanners are giant Go binaries (trufflehog, gitleaks). This is a 500-line Python file that runs as an MCP stdio server, so any LLM agent (Claude Desktop, Cursor, Windsurf, Cline) can ask it scan_directory and get a structured report in their tool-call response.

Use cases:

• Pre-commit hook in CI
• Agent-driven code review ("did this PR leak credentials?")
• Audit a freshly-cloned repo before opening it in your shell
• Inline scan during agent file edits

Install

# Python (recommended)
git clone https://github.com/eltociear/secrets-audit-mcp.git
python3 secrets-audit-mcp/server.py  # stdio MCP server

# Or via npm wrapper (TBD)
npm install -g @eltociear/secrets-audit-mcp

MCP client config

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

Tools

All return a risk score (0-100), severity bucket (CRITICAL/HIGH/MEDIUM/LOW/SAFE), and per-finding details with line numbers and redacted matches.

Coverage

Providers (32 rules total):

• Cloud: AWS access/secret, GCP API key + service-account JSON, Heroku, DigitalOcean, Cloudflare
• Source/CI: GitHub PAT/OAuth/App/Refresh/Fine-grained, npm tokens, Docker Hub PAT
• Payments: Stripe secret + restricted
• Comms: Slack bot/user/webhook, Discord bot/webhook, Telegram bot, Twilio, SendGrid, Mailgun
• AI/ML: OpenAI, Anthropic, HuggingFace, Replicate
• Web3: Ethereum private key (context-aware), Alchemy, Infura
• Keys: RSA / EC / OpenSSH / PGP / generic PEM private keys
• Generic: JWT, apikey="..." heuristic, generic secret assignments

Each match is redacted (AKIA***MPLE) before being returned, so the report itself doesn't leak the secret to the next LLM hop.

CI usage

- name: Secrets audit
  run: |
    python3 server.py <<EOF | jq -r '.result.content[0].text'
    {"jsonrpc":"2.0","id":1,"method":"tools/call",
     "params":{"name":"scan_directory","arguments":{"path":"."}}}
    EOF

A first-class GitHub Action will ship as eltociear/secrets-audit-action@v1.

Sister project — skill-audit-mcp

skill-audit-mcp covers behavioral malware patterns (download-and-execute, prompt injection, credential exfiltration). Run both for full coverage:

... View full README on GitHub