xploitscan-mcp

Model Context Protocol (MCP) server that exposes XploitScan as tools your AI coding agent can call. Scan AI-generated code for vulnerabilities from inside Claude Desktop, Cursor, Windsurf, or any other MCP client.

What this is

XploitScan is a static security scanner tuned for the kind of code AI coding tools (Cursor, Lovable, Bolt, Replit, Claude Code) produce by default — unprotected webhooks, hardcoded API keys, SQL injection via template literals, missing auth middleware, CORS misconfigurations. 210 rules, all documented at xploitscan.com/rules.

This package wraps the same scanner engine in the Model Context Protocol so an AI coding agent can check its own output before writing it to disk. Three tools:

• scan_code — scan files, return findings with rule IDs, severity, and fix suggestions
• explain_rule — look up what a specific rule means (e.g., VC005 = unprotected Stripe webhook)
• grade_code — single-call "is this safe to ship" — returns a letter grade A+ through F plus summary

Runs locally over stdio. No network calls, no API key required, no telemetry.

Install

Claude Desktop

Edit your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the xploitscan server under mcpServers:

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

Restart Claude Desktop. The three tools show up in the tool picker.

Cursor

Cursor reads MCP config from ~/.cursor/mcp.json:

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

Other MCP clients

Any client that supports the standard MCP stdio transport. Command: npx -y xploitscan-mcp. No flags.

Usage

Once wired up, just ask your agent to scan. Examples:

"Scan the code you just wrote for security issues."

"Is this safe to ship?" (agent will call grade_code)

"What does VC005 mean?" (agent will call explain_rule)

The agent decides when to call the tool. If you want explicit invocation, tools can usually be triggered directly from the client's tool picker.

Tool schemas

scan_code

Input:

{
  "files": [
    { "path": "src/api/login.ts", "content": "..." },
    { "path": "src/db/users.ts", "content": "..." }
  ]
}

Output: text summary + structured { findings, filesScanned, findingCount }.

explain_rule

Input:

{ "ruleId": "VC006" }

Output: rule title, severity, category, OWASP + CWE mapping, description.

grade_code

Input: same as scan_code.

Output: { grade, score, summary, findingCount, breakdown, findings }.

Relationship to the XploitScan CLI

The MCP server runs the same scanner engine as the XploitScan CLI (npm i -g xploitscan) and the web scanner at xploitscan.com/scan. Findings are identical across all three surfaces.

For CI, use the GitHub Action. For one-off local scans from the shell, use the CLI. Use this MCP server when you want your AI coding agent to self-check its output.

Privacy

All analysis runs locally in the subprocess started by your MCP client. Code never leaves your machine. No telemetry. No account required.

License

MIT.