Mostly yes — they were written by Claude/Gemini/ChatGPT, but they've already caught several regressions introduced by the models during later refactors; there’s a fascinating loop in letting one iteration of an LLM audit the next via a static test suite.
The architecture is moving in that direction — v2.0 is the idea of a multi-source system. Not there yet and my take a while, but the concept is documented if you want to look at the approach.
AnythingLLM + qwen2.5:14b is what I use — good for pattern recognition across metrics you wouldn't think to look for manually.
For everything else the HTML dashboards honestly do a better job. garmin_analysis.html shows your daily values, personal 90-day baseline and age/fitness reference ranges all in one view. Most questions answer themselves just by looking at it.
I wanted to ask an AI questions about my Garmin health data — sleep, HRV, stress, recovery — without sending it to OpenAI or any other cloud service.
Built this with Claude as my coding partner. I can't write Python. It works anyway.
What it does: downloads your complete Garmin Connect data locally, exports to Excel and interactive HTML dashboards, generates RAG-optimised JSON summaries for Ollama/AnythingLLM, and includes a Windows desktop app so no terminal is needed.
Everything runs on your own machine. Nothing leaves it.
GitHub: github.com/Wewoc/garmin_collector