security-mcp

MCP server for security screening: prompt-injection defense, PII detection, and secrets scanning.

This crate is intended to sit “in front of” other tools/servers, so inputs and outputs can be screened consistently.

Status

Alpha / under active development. Rules and thresholds will evolve.

What It Does Today (Code-Backed)

• Runs an MCP-compatible JSON-RPC server over HTTP/WebSocket.
• Screens text in two directions:
  • Input screening (prompt-injection patterns, suspicious encodings)
  • Output screening (PII/secrets patterns, high-entropy tokens)
• Supports batch scanning with CPU-parallelism (rayon).

What It Does Not Do Yet

• Full DLP-grade PII detection (this is heuristic/pattern-based today).
• Policy-driven redaction across structured formats beyond plain text.

Running

cargo run -p security-mcp -- --help

Usage with VS Code / GitHub Copilot

Install the binary:

cargo install security-mcp

Add to your VS Code MCP configuration (typically ~/.config/Code/User/profiles/<profile>/mcp.json or .vscode/mcp.json):

{
  "servers": {
    "security-mcp": {
      "type": "stdio",
      "command": "security-mcp",
      "args": ["--stdio"]
    }
  }
}

Important: The --stdio flag is required for VS Code integration. Without it, the server defaults to HTTP mode on port 3001.

Usage with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "security-mcp": {
      "command": "security-mcp",
      "args": ["--stdio"]
    }
  }
}

License

MIT