llama

This MCP server acts as a bridge between Claude Desktop and your local LLM instance exposed via llama-server. It enables LibreModel-style conversations to flow through the MCP protocol, giving you full control over model parameters such as temperature, max_tokens, top_p, and top_k, while providing health checks, testing tools, and usage metrics. To use it, ensure llama-server is running and that Claude Desktop is configured to connect to the MCP server defined in your environment. The server exposes a chat tool for ongoing conversations, a quick_test tool for capability checks, and a health_check tool to monitor status. Once configured, Claude Desktop can send MCP messages to the server, which then translates them into llama-server API calls and returns responses back to Claude Desktop.

Prerequisites:

• Node.js installed (recommended LTS version)
• Access to a running llama-server instance (with your model loaded)
• Claude Desktop configured to point at the MCP server (per environment setup)

1. Install the MCP server package

• Run in your project directory:

npm install @openconstruct/llama-mcp-server

2. Build or prepare the server (if you build step is required by the package)

npm run build

3. Run the llama MCP server

• Start the server (adjust path as needed to your built index):

node dist/index.js

4. Verify llama-server is up and accessible at the configured URL

• Ensure llama-server is running with your model, for example:

./llama-server -m lm37.gguf -c 2048 --port 8080

5. Configure Claude Desktop to use the MCP server

• Edit ~/.config/claude/claude_desktop_config.json and ensure you have:

{
  "mcpServers": {
    "libremodel": {
      "command": "node",
      "args": ["/home/jerr/llama-mcp/dist/index.js"]
    }
  }
}

6. Restart Claude Desktop and test the connection using the built-in tools (chat, quick_test, health_check).

Notes and tips:

• Environment variable: LLAMA_SERVER_URL should point to your llama-server API (default http://localhost:8080).
• If you see "Cannot reach LLama server", verify that llama-server is running and that the port matches LLAMA_SERVER_URL. Check firewall rules.
• After changing configuration, restart Claude Desktop to ensure MCP server changes are picked up.
• The npm package name to install is @openconstruct/llama-mcp-server. The server in this setup is configured under the libremodel key; you can rename the MCP server entry if you want multiple instances.
• For troubleshooting, ensure the dist/index.js path used in the Node command is the correct built artifact and that permissions allow execution.