Multi MCP

A flexible and dynamic Multi-MCP Proxy Server that acts as a single MCP server while connecting to and routing between multiple backend MCP servers over STDIO or SSE.

🚀 Features

• ✅ Supports both STDIO and SSE transports

• ✅ Can connect to MCP servers running in either STDIO or SSE mode

• ✅ Proxies requests to multiple MCP servers

• ✅ Automatically initializes capabilities (tools, prompts, resources) from connected servers

• ✅ Dynamically add/remove MCP servers at runtime (via HTTP API)

• ✅ Supports tools with the same name on different servers (using namespacing)

• ✅ Deployable on Kubernetes, exposing a single port to access all connected MCP servers through the proxy

📦 Installation

To get started with this project locally:

🖥️ Running Locally

You can run the proxy locally in either STDIO or SSE mode depending on your needs:

1. STDIO Mode

For CLI-style operation (pipe-based communication). Used for chaining locally executed tools or agents.

2. SSE Mode

Runs an HTTP SSE server that exposes a /sse endpoint. Useful for remote access, browser agents, and network-based tools.

Note: You can also configure the host and port using --host / --port arguments.

3. Production Mode (with External MCP Servers)

For production deployments with external MCP servers (GitHub, Brave Search, Context7), use the included startup script:

The production configuration is stored in msc/mcp.json (git-ignored for security). This configuration includes:

• GitHub MCP Server: Repository management, issues, pull requests

• Brave Search MCP Server: Web search capabilities

• Context7 MCP Server: Library documentation and code examples

All servers use environment variable interpolation for secrets (e.g., ${GITHUB_PERSONAL_ACCESS_TOKEN}).

⚙️ Configuration

The proxy is initialized using a JSON config (default: ./mcp.json):

This config defines the initial list of MCP-compatible servers to spawn and connect at startup.

Note: Tool names are namespaced internally as server_name::tool_name to avoid conflicts and allow multiple servers to expose tools with the same base name. For example, if an MCP server named calculator provides an add tool, it will be referenced as calculator::add.

You can also connect to a remote MCP server using SSE:

More examples can be found in the examples/config/ directory.

🔄 Dynamic Server Management (SSE only)

When running in SSE mode, you can add/remove/list MCP servers at runtime via HTTP endpoints:

Example to add a new server:

add_server.json:

🐳 Docker

You can containerize and run the SSE server in K8s:

Kubernetes

You can deploy the proxy in a Kubernetes cluster using the provided manifests.

Run with Kind

To run the proxy locally using Kind:

Exposing the Proxy

The K8s manifest exposes the SSE server via a NodePort (30080 by default): You can then connect to the SSE endpoint from outside the cluster:

Connecting to MCP Clients

Once the proxy is running, you can connect to it using any MCP-compatible client — such as a LangGraph agent or custom MCP client.

For example, using the langchain_mcp_adapters client, you can integrate directly with LangGraph to access tools from one or more backend MCP servers.

See examples/connect_langgraph_client.py for a working integration example.

Make sure your environment is set up with:

• An MCP-compatible client (e.g. LangGraph)

• .env file containing:

Inspiration

This project is inspired by and builds on ideas from two excellent open-source MCP projects:

• mcp-proxy by @sparfenyuk

• fastmcp by @jlowin