Cut — AI calorie & macro tracker

A minimal, liquid-glass calorie tracker built for cutting. Snap a photo of your food, an AI reads the plate, and your day's calories + macros update automatically. Black-and-white UI, pastel charts, full dark mode.

• 📸 Snap & log — photograph a meal; Gemini 2.5 Pro estimates each item's calories, protein, carbs, fat, fiber, sugar & sodium, with a confidence score per item. Low-confidence items get flagged before they're added.
• 💬 Coach that learns — correct the AI in plain English ("that's 1 cup of rice, not 2"). Corrections are saved and fed back into future prompts, so it gets more accurate the more you use it.
• ⌨️ Manual AI entry — just type "1 cup rice, 6 oz chicken" and it logs the macros.
• 🎯 Smart onboarding — height/weight/activity/goal → Mifflin-St Jeor TDEE, a recommended calorie deficit, and protein/carb/fat/fiber targets, with safety warnings if your pace is too aggressive.
• 📊 Progress — pastel charts for calories, protein, fiber and weight trend, plus adherence and weight-change stats.
• 🍽️ Meals — food is grouped into breakfast / lunch / dinner / snacks, each with its own calorie subtotal. The right meal is auto-picked by time of day.
• ⚡ Quick add (no AI) — re-log foods you eat often with one tap, straight from a "recent foods" row — saves your limited API calls.
• 📅 History — step back to any past day to review, edit, or back-fill it.
• 🤖 "What should I eat?" — AI suggests a meal that fits your remaining macros for the day, prioritising protein.
• 🔥 Streaks — a logging streak to keep you consistent.
• 📱 Installable (PWA) — add it to your phone's home screen and it runs full-screen like a native app, offline shell included.
• 🔄 Cloud sync — log on your phone, review on your laptop. Everything's in Postgres.
• 💧 Extras — water tracking, daily reset, inline-editable items, weight logging.

• BMR: Mifflin-St Jeor — the most accurate predictive equation for adults.
• TDEE: BMR × activity factor (1.2–1.9).
• Deficit: ~275 / 550 / 825 kcal/day for relaxed / steady / aggressive (1 kg fat ≈ 7700 kcal). Floored at 1500 (men) / 1200 (women) kcal.
• Protein: 2.0 g/kg bodyweight to preserve muscle on a deficit (ISSN).
• Fat: 0.8 g/kg (hormone floor). Carbs: fill the rest. Fiber: 14 g / 1000 kcal.

npm install

Create one at https://aistudio.google.com/apikey.

Any Postgres works. Easiest free options:

• Neon — https://neon.tech → create project → copy the connection string.
• Supabase — https://supabase.com → Project → Settings → Database → URI.
• Vercel Postgres — add it from the Vercel dashboard (auto-sets the env var).

A local Postgres also works for development — the schema is created automatically on first run. But for phone↔laptop sync you want a hosted DB.

Copy the template and fill it in:

cp .env.example .env.local

GEMINI_API_KEY=AIza...             # optional — you can also paste it in-app
GEMINI_MODEL=gemini-2.5-flash      # free-tier friendly; swap to pro/newer anytime
DATABASE_URL=postgres://user:pass@host/db?sslmode=require

Tip: the Gemini key is optional in the env file. You can link it from Profile → AI connection instead — paste your key, hit Link, and it auto-runs a connection test. A key set there is stored in your database (so it syncs across devices) and overrides the env var. The status dot tells you at a glance whether a working key is linked.

npm run dev

Open http://localhost:3000 — you'll be taken through onboarding first.

1. Push this folder to a Git repo and import it in Vercel (or use the vercel CLI).
2. In Project → Settings → Environment Variables, add GEMINI_API_KEY, GEMINI_MODEL, and DATABASE_URL.
3. Deploy. The database schema auto-creates on the first request.

Since it's just for you, there's no auth — all data lives under a single user.

Sign-in is powered by Supabase Auth. When NEXT_PUBLIC_SUPABASE_URL and NEXT_PUBLIC_SUPABASE_PUBLISHABLE_KEY are set, the app requires sign-in and scopes every row to the logged-in user's id, so you can log in from any device and see the same data. If those vars are absent (e.g. pure local dev), the app runs in single-user mode with no login.

Two sign-in methods are built in:

• Email magic link — works out of the box, no provider setup.
• Continue with Google — needs Google enabled in Supabase (steps below).

1. Supabase → Authentication → Sign In / Providers → Google → enable.
2. Create OAuth credentials in Google Cloud Console → APIs & Services → Credentials → OAuth client ID → Web application.
   • Authorized redirect URI: https://<your-project>.supabase.co/auth/v1/callback (Supabase shows the exact URL on the Google provider page).
3. Paste the Client ID and Client secret into Supabase → Save.
4. Supabase → Authentication → URL Configuration:
   • Site URL: your production URL (e.g. https://cut-eta.vercel.app).
   • Redirect URLs: add https://cut-eta.vercel.app/auth/callback (and http://localhost:3000/auth/callback for local testing).

That's it — /auth/callback exchanges the code for a session, and the proxy keeps it refreshed. Sign out from Profile → Sign out.

Data security: queries go through the pooled Postgres connection and are filtered by the server-verified user id (not anything the client sends), so users only ever see their own rows.

Everything routes through env vars in lib/gemini.ts:

• GEMINI_MODEL — used for image analysis (default gemini-2.5-flash).
• GEMINI_TEXT_MODEL — optional, used for text chat/manual entry. Defaults to GEMINI_MODEL. Set it to gemini-2.5-flash to make typing faster/cheaper.

When Google ships a newer model, just change GEMINI_MODEL — no code changes.

Next.js 16 (App Router) · React 19 · TypeScript · Tailwind v4 · Recharts · postgres · @google/genai.