A Model Context Protocol (MCP) multiplexer that intelligently routes to the right tool from hundreds of available options using semantic search.
Problem: Managing multiple MCP servers means dealing with hundreds of tools, making it hard for LLMs to find the right one.
Solution: mcp-mux acts as an intelligent router that exposes just two tools:
search- Find tools using natural languageexecute- Run the tool you found
This dramatically simplifies tool discovery and improves LLM performance with MCP servers.
- Python 3.10+
- MCP-compatible client (e.g., Claude Desktop)
# Using uvx (recommended)
uvx mcp-mux init
# Or install globally
pip install mcp-mux- Initialize configuration:
uvx mcp-mux init- Configure your MCP servers in
~/.mux/config.json:
{
"mcpServers": {
"filesystem": {
"command": "uvx",
"args": ["mcp-server-filesystem", "/Users/me/projects"]
},
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"]
}
}
}- Add mcp-mux to your MCP client configuration:
{
"mcpServers": {
"mux": {
"command": "uvx",
"args": ["mcp-mux"]
}
}
}Instead of exposing hundreds of tools from multiple servers, mcp-mux exposes just two:
search- Find tools using natural languageexecute- Run a specific tool
# List available embedding models
uvx mcp-mux model --list
# Switch to a different model
uvx mcp-mux model BAAI/bge-small-en-v1.5MIT
This project is in early development. Features may change.