Model Context Protocol (MCP)

Yumi supports the Model Context Protocol (MCP), a standard created to let AI tools safely expose local capabilities (like terminal commands or system states) to third-party LLM clients.

By activating her MCP bridge, you can connect Yumi's registry of system tools directly into apps like Claude Desktop or Cursor.

🛠️ Step 1: Initialize Yumi's MCP Endpoint

Yumi implements an MCP transport layer in her backend package under src/yumi/mcp/.

To launch Yumi solely as a background MCP server (without her web avatar), execute the CLI command:

uv run yumi mcp start

This establishes an MCP server pipeline listening on standard inputs/outputs (stdio), ready to receive tool executions.

💻 Step 2: Configure Claude Desktop Integration

To let your desktop AI chat utilize Yumi's tools, register her in your local Claude config file:

File Path: C:\Users\YourName\AppData\Roaming\Claude\claude_desktop_config.json
Linux/macOS Path: ~/Library/Application Support/Claude/claude_desktop_config.json

Add Yumi to your "mcpServers" object:

{
  "mcpServers": {
    "yumi-companion": {
      "command": "uv",
      "args": [
        "--directory",
        "C:\\Users\\YourName\\Yumi",
        "run",
        "yumi",
        "mcp",
        "start"
      ]
    }
  }
}

🚀 Step 3: Test the Connection

Restart Claude Desktop.
You will see a small hammer icon indicating that MCP tools are loaded.
Type: "What time is it on my computer?"
Claude will call Yumi's get_current_time tool securely in the background, return the output, and display the result.

Proceed to the Ops & Reference section to troubleshoot and optimize performance!