Building my own health coach

I didn’t want to pay for Google Health Coach to explain data my watch was already collecting. And the questions I actually ask aren’t only about the watch. Is something missing from my gym routine? Why have my lifts stopped moving? Does the food I usually buy support the training I’m doing? I already had an agent to talk to; I needed to give it records that could answer those questions.

The watch supplied the first part. I started saving data from Google’s account-level Health API into my own database: sleep, heart rate, activity and other measurements when they’re available. A scheduled collector goes back over recent days because Google’s figures can change after the first fetch. That gives me a history to compare, not just today’s summary. But the watch can tell me I was active without knowing what I actually lifted.

My gym app records the sets, reps and weights I finish. The health system can also detect a gym visit using presence and heart-rate data. I keep those facts separate: turning up isn’t the same as completing a session, and forgetting to log a session isn’t proof that I stayed home. Together, they let the agent check both attendance and progress without inventing the missing half.

Food was the other missing piece. I do most of my shopping at one supermarket, and the itemised receipts are collected automatically. Purchase history and product nutrition information give an estimate of weekly at-home macros, so I don’t enter each meal by hand. Nor can a receipt prove what I ate: food from elsewhere, leftovers, sharing and unknown nutrition labels all leave gaps. That’s important when the question is whether I might be short on protein or fibre, rather than simply what I bought.

The useful conversations are already quite specific. I asked what was missing from my lower-body routine; the agent checked the exercises I had logged rather than suggesting a generic workout. When progress felt stuck, we looked at lift history alongside my weight trend instead of assuming the answer was more gym visits. And when I asked where I could cut calories, the recurring grocery purchases offered concrete options in my usual shopping rather than a stock meal plan. I could make a change without keeping a manual food diary.

The databases still belong to their separate apps. The agent reads them through MCP tools and brings the relevant evidence together when I ask; it isn’t one giant health record. That makes the limits visible too. When I once asked whether lifting affected my sleep, the reliable overlap was too small to give a confident answer. I’d rather hear that than a neat story drawn from thin data. Maintaining the collectors and integrations is the trade-off for skipping Google’s subscription. What I get back is a coach that can look beyond the watch—and show me what it actually knows.