mcp-routing-gateway
An MCP routing proxy that prevents LLMs from executing unintended tools. It securely curates, filters, and routes tools from multiple backend MCP servers. (LLMが意図しないツールを実行するのを防ぐMCPルーティングプロキシ。複数のバックエンドMCPサーバーからのツールを安全にキュレーション、フィルタリング、ルーティングします。)
README
MCP Routing Gateway
The MCP Routing Gateway is a stateless router and Pure Proxy designed to solve the problem of "LLMs selecting tools unintended by the user," which is a common issue faced by current mainstream AI agents (e.g., Claude Desktop, Cline).
🌟 Why MCP Routing Gateway?
When multiple MCP servers are registered directly to a standard AI agent, the LLM can freely browse and execute all tools, posing the risk of triggering unexpected destructive operations or unnecessary tool calls.
This project sits between the AI agent and the actual group of MCP servers, mediating the communication to provide a control layer that "presents only safely curated (filtered and virtualized) tools to the LLM."
🏗️ Architecture
This project incorporates both "MCP Server" and "MCP Client" functionalities internally.
AI Agent (MCP Client) → | [MCP Server] Gateway [MCP Client] | → Backend MCP Servers
- Frontend (MCP Server): Acts as a "single, secure MCP server" to the AI agent, accepting connections via standard input/output (
stdio). - Backend (MCP Client): Reads
mcp_config.jsonand automatically launches and manages multiple MCP servers defined within it (viastdiospawned processes).
🎯 Design Philosophy
This project is built on the strict philosophy of being a "Router and Pure Proxy."
- Zero Payload Interference (Pass-through): It does not interfere with payloads such as request IDs or arguments. It focuses entirely on relaying communication, functioning as a pure pipe.
- Facade Pattern (Filtering and Tool Overrides):
The Gateway only intervenes in the routing of
tools/list(presenting the tool list) andtools/call(tool invocation). Based on the configuration file, it hides dangerous tools or replaces them with secure virtual tools.
⚙️ Configuration
All configurations are entirely managed in JSON format. By default, the gateway looks for these files in the working directory ~/.mcp-gateway. You can change this directory using the --work-dir CLI option.
mcp_config.json(Backend Definition): The standard configuration file format for MCP clients. It defines the group of MCP servers that the Gateway connects to and launches in the background.gateway_config.json(Routing and Filter Definition): A rule configuration file unique to the Gateway. It defines "which tools to hide (Blocked Tools)" and "which tools to virtualize (Virtual Tools)."
For detailed usage and CLI options, please see the User Manual.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。