mcp-gateway
MCP Gateway
MCP 生态的安全中间层 — 在 LLM 与工具服务器之间建立防护屏障
它解决什么问题?
当 LLM Agent 通过 MCP 协议调用外部工具时,存在三个核心风险:
凭据泄露 — 工具返回的响应中可能包含 API Key、Token 等敏感信息,直接暴露给 LLM 上下文
隐私数据外泄 — 用户个人信息(姓名、身份证、银行卡号)可能在工具调用链中被传递
恶意工具注入 — 工具描述中可能隐藏 prompt injection 指令,诱导 Agent 执行危险操作
MCP Gateway 作为代理层拦截所有流量,在请求/响应到达 Agent 之前进行安全过滤。
Related MCP server: @guardbee/mcp-security-proxy
架构概览
┌─────────────┐ ┌──────────────────────────────────┐ ┌─────────────┐
│ │ │ MCP Gateway │ │ │
│ LLM Agent │ ───► │ ┌──────────┐ ┌──────────────┐ │ ───► │ MCP Server │
│ │ │ │ Sanitize │ │ Plugin │ │ │ (tools) │
│ │ ◄─── │ │ Request │ │ Pipeline │ │ ◄─── │ │
└─────────────┘ │ └──────────┘ └──────────────┘ │ └─────────────┘
│ ▲ │ │
│ │ ┌──────────▼──┐ │
│ │ │ Sanitize │ │
│ │ │ Response │ │
│ │ └─────────────┘ │
└──────────────────────────────────┘核心流程:
请求方向:Plugin Pipeline 对参数进行脱敏(如移除 PII、过滤注入指令)
响应方向:对工具返回值进行 Token 掩码、敏感信息过滤
启动阶段:Security Scanner 对所有配置的 MCP Server 进行信誉评估
快速开始
安装
git clone <your-repo-url>
cd mcp-gateway
pip install -e .可选依赖:
pip install -e .[presidio] # 启用 PII 检测(基于 Microsoft Presidio)最简配置
在项目根目录创建 mcp.json:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
}
}
}启动
# 启用基础 Token 掩码
mcp-gateway -p basic
# 启用 Token 掩码 + PII 检测
mcp-gateway -p basic -p presidio
# 调试模式
LOGLEVEL=DEBUG mcp-gateway -p basic集成到 Cursor / Claude Desktop
{
"mcpServers": {
"mcp-gateway": {
"command": "mcp-gateway",
"args": [
"--mcp-json-path", "~/.cursor/mcp.json",
"-p", "basic",
"-p", "xetrack"
],
"servers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
}
}
}
}
}{
"mcpServers": {
"mcp-gateway": {
"command": "<python-path>",
"args": [
"-m", "mcp_gateway.server",
"--mcp-json-path", "<path-to-config>",
"-p", "basic"
],
"servers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
}
}
}
}
}安全防护能力
Token 掩码 (basic 插件)
自动识别并替换响应中的敏感凭据,支持 12 种主流云平台和开发工具的密钥格式:
类型 | 示例格式 |
AWS Access Key |
|
GitHub Token |
|
Azure Client Secret |
|
GCP API Key |
|
JWT Token |
|
HuggingFace Token |
|
GitLab Session Cookie |
|
Slack App Token |
|
Microsoft Teams Webhook |
|
mcp-gateway -p basicPII 检测 (presidio 插件)
基于 Microsoft Presidio 引擎,自动识别并匿名化文本中的个人身份信息:
信用卡号、IP 地址、电子邮箱
电话号码、身份证号(SSN)
更多实体类型见 Presidio 文档
pip install -e .[presidio]
mcp-gateway -p presidio安全扫描器 (--scan)
启动前对所有 MCP Server 进行信誉评估和工具描述分析:
mcp-gateway --scan -p basic扫描维度:
信誉评估 — 基于 GitHub 数据(Star、Fork、Issue 活跃度)和 NPM 下载量计算综合评分
工具描述扫描 — 检测隐藏的 prompt injection 指令、敏感文件路径引用、危险操作指令
自动阻断 — 信誉分低于阈值(默认 30 分)的 Server 会被标记为
blocked并阻止加载
扫描结果会写入配置文件:
{
"servers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."],
"blocked": "passed"
}
}
}状态值:"passed" (安全) | "blocked" (阻断) | "skipped" (手动跳过) | null (未扫描)
调用追踪
Xetrack 追踪插件
记录所有工具调用的完整上下文,支持 SQLite 和 DuckDB 查询:
pip install xetrack
mcp-gateway -p xetrack环境变量配置:
XETRACK_DB_PATH— SQLite 数据库路径XETRACK_LOGS_PATH— 日志文件目录
{
"mcpServers": {
"mcp-gateway": {
"command": "mcp-gateway",
"args": ["--mcp-json-path", "~/.cursor/mcp.json", "-p", "xetrack"],
"env": {
"XETRACK_DB_PATH": "tracing.db",
"XETRACK_LOGS_PATH": "logs/"
}
}
}
}查询示例:
from xetrack import Reader
df = Reader("tracing.db").to_df()-- DuckDB
INSTALL sqlite; LOAD sqlite; ATTACH 'tracing.db' (TYPE sqlite);
SELECT server_name, capability_name, content_text FROM db.events LIMIT 10;代理工具
Gateway 向 LLM 暴露两个标准化工具:
工具 | 功能 |
| 获取所有已注册 MCP Server 的能力列表,帮助 LLM 选择合适的工具 |
| 通过 Gateway 执行任意 MCP 工具调用,自动进行请求/响应的安全处理 |
插件开发
插件系统基于 ABC 基类 + 装饰器注册模式:
from mcp_gateway.plugins.base import GuardrailPlugin
from mcp_gateway.plugins.manager import register_plugin
@register_plugin
class MyPlugin(GuardrailPlugin):
@property
def name(self) -> str:
return "my-plugin"
def process_request(self, context):
# 请求方向的处理逻辑
return context.arguments
def process_response(self, context, response):
# 响应方向的处理逻辑
return response插件通过 PluginManager 自动发现和加载,支持请求/响应双向拦截。
项目结构
mcp_gateway/
├── __init__.py # 包入口
├── server.py # MCP Server 生命周期管理
├── gateway.py # 动态工具注册、CLI 参数解析
├── config.py # 配置文件加载
├── sanitizers.py # 请求/响应安全分发
├── plugins/
│ ├── base.py # Plugin ABC 基类
│ ├── manager.py # 插件发现、注册、Pipeline
│ ├── guardrails/
│ │ ├── basic.py # Token 掩码插件
│ │ └── presidio.py # PII 检测插件
│ └── tracing/
│ └── xetrack.py # 调用追踪插件
├── security_scanner/
│ ├── scanner.py # 扫描器主入口
│ ├── github_collector.py # GitHub API 数据采集
│ ├── npm_collector.py # NPM Registry 数据采集
│ ├── smithery_collector.py# Smithery 市场数据采集
│ ├── project_analyzer.py # 综合信誉评分算法
│ └── tool_poisoning_analyzer.py # 工具描述安全分析
└── tests/
├── test_sanitizers.py
├── test_tool_poisoning_analyzer.py
├── test_plugin_pipeline.py
└── test_config.pyLicense
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Security & DLP proxy for MCP: tool-poisoning scans, PII redaction on tool args/results. Beta.
Security firewall for AI agents — scans MCP calls for injection, secrets, and risks.
Safety tools for AI agents: safe page reader, prompt-injection, secrets, PII and MCP server scans
Real-time prompt injection filtering, MCP payload inspection, and A2A trust scoring.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables safe use of any MCP server by proxying and live-scanning all tool requests and responses, blocking or redacting poison descriptions, indirect prompt injection, malicious arguments, and unauthorized destinations.-
- AlicenseNot gradedqualityFmaintenanceEnables transparent security for any MCP server by intercepting tool calls, blocking prompt injection attempts, masking PII in responses, and writing immutable audit logs.MIT
- AlicenseNot gradedqualityBmaintenanceEnforces MCP security by proxying between AI agents and MCP servers, scanning tools and results for prompt injection, enforcing allow/deny policies, redacting sensitive arguments, and logging all traffic.56 PyPIMIT
- AlicenseNot gradedqualityCmaintenanceEnables secure, low-latency proxying of MCP tool calls by detecting and blocking malicious requests such as path traversals and command injections, while redacting leaked credentials.MIT