MCP Guardrail Server

Read this in other languages: 日本語

MCP Guardrail Server is a secure MCP (Model Context Protocol) server that executes only pre-authorized commands. It can be used in high-security environments to provide AI assistants with limited command execution capabilities.

Features

• Executes only commands included in the allowlist
• Command execution timeout functionality

Installation

npm install
npm run build

Usage

npm start -- [--allowed-commands <comma-separated-list>]

Options

• --allowed-commands: Comma-separated list of allowed Bash commands (optional, default: git,ls,mkdir,cd,npm,npx,python)

Examples

# Run in development mode
npm run dev

# Start server (using default allowed commands)
npm start

# Start with custom command list
npm start -- --allowed-commands git,ls,node

Configuration with Claude Desktop

To use this MCP server with Claude Desktop, add the following entry to your Claude Desktop configuration file (typically ~/.config/Claude Desktop/claude_desktop_config.json):

{
  "mcpServers": {
    "guardrail": {
      "command": "node",
      "args": [
        "/absolute/path/to/dist/index.js",
        "--allowed-commands",
        "git,ls,node,echo" // Add your allowed commands here
      ],
      "env": {}
    }
  }
}

Connectivity Test

To verify that the server is working correctly, run:

npm test

This will run the following tests:

• Executing an allowed command
• Attempting to execute an unauthorized command (which should be rejected)

Security Notes

• Only allow commands that are absolutely necessary
• Do not allow potentially dangerous commands (such as rm -rf)

License

MIT