Run local Ollama models as an MCP server so an agent can send work to a model on your own machine instead of an API.