I built Metrify because every tracker I tried made logging a four-item dinner take two minutes of searching, scrolling, and setting serving sizes. I'd quit by week three every time.

Then I noticed I'd started just telling ChatGPT what I ate, and a lot of other people were doing the same thing. It works great for about a day. Then it's gone. There's no persistent log, nothing to look back at, and the numbers aren't as reliable as something like MyFitnessPal.

So I built the thing in the middle. The whole app is a conversation, but it's reading from and then writing to a real database.

"Chicken breast, rice, broccoli, and a Diet Coke" logs four entries from one message. "Actually make that two eggs, not three" edits the row it just wrote instead of adding a duplicate. "Same lunch as yesterday" pulls the real entry. You can log a workout, a weigh-in, or last night's sleep in the same thread.

Everything also has a manual option, same as any traditional tracking app. The agent can get you 90% of the way there, then you can very easily edit a serving size or add/replace an item if needed,

The part I'm most proud of: the assistant has 46 tools with direct access to the local database, so it can write, edit, delete, and query your actual data. It's not an advice layer bolted onto a normal app. But the model never does arithmetic. It parses what you said and picks the right food out of a 447,000-item USDA-backed catalog, then fixed formulas on your device compute every calorie, macro, and projection.

I also built a test harness that runs 300+ queries against the assistant, plus some really cool retrieval and prompt structure. That work is why it runs on a genuinely cheap model without giving up accuracy (still a work in progress, but it's genuinely pretty good right now, shows consistently over 90% accuracy in my intentionally difficult testing sandbox), and it's the only reason I can afford to give every user 3 free AI chats a day, forever.

I also built out a detailed analytics page, with predictive weight gain/loss, various insights, and an assortment of graphs based on the data you log. All of this is accessible to the agent as well, so you can ask any questions or receive personalized advice based on what you have logged, making the app increasingly useful as you use it.

Two other choices worth mentioning. Everything lives in SQLite on your phone. No account, no server holding your food/exercise logs. Plus, your calorie goal doesn't go up when you log a workout, because a sedentary baseline plus logged burn counts your training once instead of twice (see more about that on the website under articles on why that's useful).

Syncs weight, workouts, steps, and sleep through Apple Health and Google Health Connect. The agent also has full access there so you can either do it yourself, set it to automatic mode, or simply say "log my steps".

Happy to answer anything, especially about the tool-calling architecture or the on-device tradeoffs. Any feedback is really appreciated!

Metrify

Fully agentic fitness app backed by a USDA dataset

Fully agentic fitness app backed by a USDA dataset