The photo is sent to Claude, which names each food, estimates its weight, and returns calories and macros per item. The photo is analysed and thrown away — nothing is stored on the server.
What it's good at
Identifying food. Chicken thigh vs. breast, jasmine rice vs. quinoa, which cut of steak — that part is reliable.
Where it's weakest
- Portion size. One flat photo has no depth. Expect 20–40% error unless something in frame gives scale. Shoot at a 45° angle with your fork or hand visible.
- Invisible fat. Grilled chicken and chicken cooked in two tablespoons of oil look identical and are about 240 calories apart. The estimate always adds a separate line for cooking fat so you can see it and correct it — if you know it was dry, delete that line.
- Mixed dishes. Stews, curries, smoothies and shakes hide what's in them. Those come back marked low. Use Quick add instead.
The protocol that actually works: weigh your protein source on a kitchen scale — ten seconds — and photograph the rest. That gets you to roughly ±10% on the number that drives muscle growth, instead of ±30% on everything.
Packaged food? Point the camera at the nutrition label instead of the food. It reads the panel and beats any visual estimate.
Your targets
—
Targets live in nutri_config.php on the server. Change them there and this page picks them up — no need to touch the app.