AI导哈

MCP Guardrails 代理

一款策略驱动的MCP服务器防护代理，前置拦截PII泄露、密钥暴露和提示注入，为MCP客户端与服务器之间的交互提供安全保护。

  • by interactive-inc
  • TypeScript
  • 本地部署
  • 1 星标
  • 更新于
在 GitHub 查看 interactive-inc/open-mcp-guardrails

使用说明

日本語

A policy-based guardrails proxy that sits in front of any MCP server.

Just prepend open-mcp-guardrails to your existing MCP server command to protect your app from PII leaks, secret exposure, and prompt injection.

Requirements

Node.js 23.6+ is required because guardrails.config.ts is loaded via Node's native Type Stripping — no separate compile step needed.

Install

npm install open-mcp-guardrails

Quick Start

1. Create a Config File

npx open-mcp-guardrails init

An interactive prompt lets you choose between TypeScript and JSON:

  Select config format:
    1) TypeScript  (guardrails.config.ts)
    2) JSON        (guardrails.json)

  Choice [1]:

Use --json to skip the prompt and generate JSON directly.

TypeScript

// guardrails.config.ts
import { defineConfig, pii, secrets } from "open-mcp-guardrails";

export default defineConfig({
  rules: [
    pii().block(),
    secrets().block(),
  ],
});

Calling defineConfig() with no arguments enables PII + secret protection by default.

JSON

{
  "$schema": "https://unpkg.com/@interactive-inc/open-mcp-guardrails@latest/dist/guardrails.schema.json",
  "rules": [
    { "type": "pii", "action": "block" },
    { "type": "secrets", "action": "block" }
  ]
}

JSON covers ~90% of use cases. Use TypeScript when you need custom() rules or tool().check() with complex logic.

2. Register with Your MCP Client

Claude Code — Create .mcp.json at your project root:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "open-mcp-guardrails",
        "--",
        "npx", "@modelcontextprotocol/server-filesystem", "/tmp"
      ]
    }
  }
}

guardrails.config.ts in the current directory is auto-discovered — no -c flag needed.

Claude Desktop — Place config in ~/.config/open-mcp-guardrails/guardrails.config.ts, then add to claude_desktop_config.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "open-mcp-guardrails",
        "--",
        "npx", "@modelcontextprotocol/server-filesystem", "/tmp"
      ]
    }
  }
}

Everything before -- is for guardrails, everything after is the original MCP server command.

Config Resolution

When -c is not specified, config is auto-discovered in this order:

  1. ./guardrails.config.ts (current directory)
  2. ./guardrails.json (current directory)
  3. ~/.config/open-mcp-guardrails/guardrails.config.ts (XDG user config)
  4. ~/.config/open-mcp-guardrails/guardrails.json (XDG user config)

You can always override with -c <path> for explicit control.

Rules

Presets

The simplest way — select presets with protect:

export default defineConfig({
  protect: ["pii", "secrets", "prompt-injection"],
});
Preset Description
"pii" Block email, phone, credit card, SSN, etc.
"secrets" Block API keys, tokens, private keys, etc.
"prompt-injection" Block prompt injection attacks

JSON equivalent:

{
  "protect": ["pii", "secrets", "prompt-injection"]
}

Fluent Builder API

Define rules with the builder API for full control:

import {
  defineConfig, pii, secrets, promptInjection,
  contentFilter, flow, tool, custom,
} from "open-mcp-guardrails";

export default defineConfig({
  rules: [
    // Detection rules — scope limits to specific tools
    pii().scope("filesystem__read_file").block(),
    secrets().exclude("generic_secret").warn(),
    promptInjection().threshold(0.5).block(),
    contentFilter(["classified", /confidential/i]).block(),

    // Flow control
    flow("get_website").to("send_email").block(),

    // Tool argument validation
    tool("send_email")
      .check(args => !(args.to as string)?.endsWith("@company.com"))
      .block("Only @company.com addresses allowed"),

    // Custom logic
    custom("rate-limit")
      .phase("pre")
      .evaluate(ctx => {
        if (ctx.trace.toolCalls.length > 100) {
          return [{
            ruleName: "rate-limit",
            message: "Tool call limit exceeded",
            severity: "error",
          }];
        }
        return [];
      })
      .block(),
  ],
});

All builders end with a terminal method: .block() / .warn() / .log().

scope — Limit Detection to Specific Tools

Detection rules (pii, secrets, promptInjection, contentFilter) apply to all tools by default. Use .scope() to restrict to specific tools:

pii().scope("filesystem__read_file").block();
secrets().scope(/^filesystem__/).block();

Tool names use the {server}__{tool} format (e.g. filesystem__read_file, github__create_issue).

In JSON config:

{ "type": "pii", "action": "block", "scope": ["filesystem__read_file"] }
{ "type": "secrets", "action": "block", "scope": ["/^filesystem__/"] }

CLI

# Start the proxy (config auto-discovered)
open-mcp-guardrails -- npx @modelcontextprotocol/server-filesystem /tmp

# Explicit config path
open-mcp-guardrails -c custom.config.ts -- npx @modelcontextprotocol/server-github

# Scaffold a config file (interactive prompt)
open-mcp-guardrails init

# Scaffold JSON directly
open-mcp-guardrails init --json

# Validate a config file
open-mcp-guardrails check

# Print JSON Schema
open-mcp-guardrails schema

Supported Clients

Documentation

See the documentation site for detailed guides and API reference.

License

MIT

相关 MCP