Ollama does not implement MCP, so pointing a local model at Apple Health data takes one extra piece: an MCP-aware client that holds the connection and turns model output into tool calls. Goose, LM Studio, mcphost, and Open WebUI with mcpo all do this, and they share the same underlying server config.
Once wired, the model routes your question. Ask about your HRV trend and it calls get_freshness, then compare_periods, and answers from your real export: 61 ms this week, up 4.4% from 58 ms. Zero runtime network calls, so the prompt and the numbers stay on your hardware.
Full setup, including the "model describes tools but never calls them" fix: https://dub.sh/yvDDhXQ
Top comments (0)