patchpilots-mcp

PatchPilots security and accessibility agents as an MCP server. Catch supply chain attacks, security anti-patterns, and accessibility issues inside Claude Code, Cursor, and any MCP-compatible IDE.

What it does

Three tools, one server:

All return structured JSON with severity, impact, and remediation for every finding.

Install

Claude Code

// ~/.claude.json or project .claude.json
{
  "mcpServers": {
    "patchpilots": {
      "command": "npx",
      "args": ["-y", "patchpilots-mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Cursor

// .cursor/mcp.json
{
  "mcpServers": {
    "patchpilots": {
      "command": "npx",
      "args": ["-y", "patchpilots-mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

VS Code

// .vscode/mcp.json
{
  "servers": {
    "patchpilots": {
      "command": "npx",
      "args": ["-y", "patchpilots-mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

API key

The server looks for your Anthropic API key in this order:

1. ANTHROPIC_API_KEY environment variable
2. apiKey field in ~/.patchpilots.json

If you already use PatchPilots CLI, your global config works automatically.

Tools

security_scan

path:     string   — file or directory to scan
severity: string   — minimum severity: "critical" | "high" | "medium" | "low" (default: "medium")
roast:    boolean? — brutally honest commentary (default: false)
model:    string?  — Claude model (default: claude-sonnet-4-6)

Returns:

{
  "findings": [
    {
      "file": "src/auth.ts",
      "line": 42,
      "severity": "critical",
      "category": "injection",
      "cwe": "CWE-89",
      "title": "SQL injection in user lookup",
      "description": "User input concatenated into SQL query...",
      "impact": "Full database access...",
      "remediation": "Use parameterized queries..."
    }
  ],
  "riskScore": "critical",
  "summary": "Found 3 security issues..."
}

scan_dependencies

path:   string   — path to package.json or project root
model:  string?  — Claude model (default: claude-sonnet-4-6)

Returns:

{
  "risks": [
    {
      "package": "event-stream",
      "severity": "critical",
      "category": "known-vulnerability",
      "title": "Compromised package with malicious code",
      "description": "...",
      "remediation": "Remove and replace with..."
    }
  ],
  "riskScore": "high",
  "summary": "Found 2 supply chain risks...",
  "stats": {
    "totalDeps": 15,
    "totalDevDeps": 8,
    "risksFound": 2
  }
}

design_audit

path:     string   — file or directory to audit
severity: string   — minimum severity: "critical" | "high" | "medium" | "low" (default: "medium")
roast:    boolean? — brutally honest commentary (default: false)
model:    string?  — Claude model (default: claude-sonnet-4-6)

Returns:

{
  "findings": [
    {
      "file": "src/components/Button.tsx",
      "line": 15,
      "severity": "high",
      "category": "accessibility",
      "wcagRef": "WCAG 2.1 SC 4.1.2",
      "title": "div with onClick instead of button element",
      "description": "Interactive element uses div with onClick...",
      "remediation": "Replace <div onClick=...> with <button>..."
    }
  ],
  "designHealthScore": "high",
  "summary": "Found 4 accessibility issues..."
}

Fun features

Risk emojis — every scan shows a visual risk score: 🟢 none 🟡 low 🟠 medium 🔴 high 💀 critical

Clean scan badges — pass with no findings and get a README badge:

[![Reviewed by PatchPilots](https://img.shields.io/badge/reviewed%20by-PatchPilots%20🔒-blueviolet)](https://github.com/alavesa/patchpilots)

Roast mode — set roast=true on security_scan or design_audit for brutally honest, funny commentary on your findings. Same structured output, spicier language. Try asking your IDE:

"roast the accessibility of my code"

When to use what

Need the full crew? Use the PatchPilots CLI — all 8 agents in one command:

npx patchpilots audit ./src --write

Want automatic PR reviews? Add the PatchPilots GitHub Action:

- uses: alavesa/patchpilots@v1
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

All three use the same analysis. The MCP server is for real-time, conversational security and accessibility scanning while you code. The CLI and Action cover the rest.

License

MIT