wtf is ollama-mcp-host?

ollama-mcp-host solves a specific connection problem. A local AI model running through Ollama, such as Google's gemma, is capable but cannot look things up or touch files on its own. To give it that ability, you connect it to tools using MCP, or Model Context Protocol, an open standard for AI tools. The trouble is that Ollama and MCP servers were not built to talk to each other directly. This project is the translator that sits in between: you register one or more MCP servers with it, and a single Ollama model can automatically pick the right tool, call it, and answer using the real result it gets back. Everything runs locally or inside a private network, with no cloud service and no API key needed for the AI model itself, which makes it useful for offline or air-gapped environments where internet access is restricted. The README demonstrates this using two example Korean government tools, one for searching Korean law and one for generating Korean office documents, but says any standard MCP server that communicates over stdio will work the same way. A notable detail is that gemma models have a thinking mode that can cause them to reason internally and then forget to actually issue the tool call, which is a common source of confusion. This host works around that by always telling Ollama to turn thinking off, so tool calls fire reliably. Setting it up involves installing Node.js version 20 or newer, installing Ollama and pulling a tool-capable model like gemma4, cloning and building the MCP servers you want to use, installing this host through npm, and writing a small JSON configuration file that lists the model, the Ollama address, and the paths to each MCP server. Once configured, you run the host from the command line and get an interactive chat where you can type questions, use a tools command to list available tools, and watch the model call tools and answer using their real output. The project has only one runtime dependency, the official MCP SDK, and connects to Ollama using Node's built-in networking, keeping its footprint small. It works both as a command line tool and as a library you can import into your own code.