The useful shape is delegation — ask the local model the part of a task that should not go to a hosted provider, and keep the rest of the conversation where it is. Everything depends on Ollama already running with a model pulled; if `list_models` comes back empty, that is the thing to fix rather than the server.
A small MCP server that connects an MCP client to a locally running Ollama installation. It reports the models you have pulled, returns the detail for one of them, and forwards a question to a named model — so a conversation in one client can consult a local model without leaving it.
- Every downloaded Ollama model listed — `list_models`
- Detailed information about a specific model — `show_model`
- A question put to a named model and its answer returned — `ask_model`
- Nothing leaves the machine: the models it talks to are the ones running in your own Ollama installation
Ollama installed and running, with at least one model pulled — the README's example is `ollama pull llama2`. Python 3.10 or higher. The client runs it as `uvx mcp-ollama`, so `uv` needs to be present; the published package is `mcp-ollama` 0.1.3 with an `mcp-ollama` entry point over stdio. No account and no key.
One command — uvx mcp-ollama