MCP LLM

The easiest way to use mcp-llm is via npx, which downloads and runs the package on demand. You can also install it globally with npm.

npx -y mcp-llm

At minimum you must set LLM_MODEL_PROVIDER and LLM_MODEL_NAME. Provider must be one of: openai, openai-compatible, ollama, or bedrock. For OpenAI, also set OPENAI_API_KEY.

export LLM_MODEL_PROVIDER=openai
export LLM_MODEL_NAME=gpt-4o
export OPENAI_API_KEY=sk-...

Fine-tune the model behavior with optional env vars. LLM_BASE_URL lets you point to any OpenAI-compatible endpoint (e.g., a local LM Studio server). LLM_ALLOW_FILE_WRITE=true must be set explicitly to enable the generate_code_to_file tool.

# For a local OpenAI-compatible server:
export LLM_BASE_URL=http://localhost:1234/v1

# To allow writing code to disk:
export LLM_ALLOW_FILE_WRITE=true

# Tune sampling:
export LLM_TEMPERATURE=0.2

Add mcp-llm to your claude_desktop_config.json (or equivalent MCP client config). Pass all required and optional env vars in the env block.

{
  "mcpServers": {
    "mcp-llm": {
      "command": "npx",
      "args": ["-y", "mcp-llm"],
      "env": {
        "LLM_MODEL_PROVIDER": "openai",
        "LLM_MODEL_NAME": "gpt-4o",
        "OPENAI_API_KEY": "sk-your-key-here",
        "LLM_ALLOW_FILE_WRITE": "true"
      }
    }
  }
}

Restart Claude Desktop. The server exposes four tools: generate_code, generate_code_to_file, generate_documentation, and ask_question. Try asking Claude to delegate a coding task to the connected LLM.