AI Nutrition Coach The app speaks English and Français Source

Know what’s on your plate. Keep it yours.

Log a meal by photo, barcode, a few words or your voice. It lands on your day as an honest plate: food groups, grams, nutrients. No account, no scores, and nothing leaves your device unless you say so.

  • On this device · encrypted
  • Zero telemetry
  • Open source · MIT
The app’s Today screen: woven meal bands for breakfast and lunch, a coach question, and the Log button.

Breakfast

Oat flakes, yogurt · 284 kcal

Lunch

Lentils, tomato · 240 kcal

Lunch

Apple · 77 kcal

Dinner

Bare cotton until you log it

Log it the way it happened.

At the table, in a shop aisle, or from memory after the fact. One thumb, one plate. If the AI is off, offline or unsure, the meal still lands.

  1. Snapshot

    One photo of the plate is plenty. Its metadata is stripped, and it goes to the AI only after you agree, every time by default.

    A plate of lentils and tomatoLands as veg & fruit and protein

  2. Barcode

    Scan a pack and Open Food Facts returns the label’s own numbers, per gram. Unknown product? You can add it.

    Crunchy oat muesli, per gramLands as starch

  3. Type it

    Ranked matches as you type, with portion presets in plain sizes and grams when you know them.

    Plain yogurt, 125 gLands as dairy

  4. Describe

    Say it or write it: “a bowl of lentil salad and an apple”. Your words become foods and grams you can check.

    “Lentil salad and some bread”Lands as veg & fruit, protein and starch

  5. Paste a recipe

    Drop in the ingredient list, say how many servings you ate, and the plate is worked out from it.

    One serving of a lasagneLands as veg & fruit, protein, starch and dairy

The Log a plate sheet: Snapshot, Barcode, Type it and Describe, with recent foods and saved meals.

And the everyday shortcuts. Recent foods, saved meals, and copying yesterday’s breakfast or a whole day.

One honest plate. Real numbers, calmly.

Four food groups, present or absent, woven into the day. Tap the plate and the grams and nutrients are there, set against the EU reference intakes or your own daily guide. Estimates wear a “~”. Nothing turns red.

Lunch

Lentils 180 g · Tomato 80 g

Veg & fruit and protein on the plate. No starch, no dairy.

This plate’s nutrition demo day, real food data

Energy
240 kcal
about 12% of the EU reference intake (2000 kcal)
Protein
19 g
about 37% of the reference (50 g)
Carbohydrate
32 g
about 12% of the reference (260 g)
of which sugars
2.9 g
about 3% of the reference (90 g)
Fat
1 g
about 1% of the reference (70 g)
Fibre
16 g
about 64% of a 25 g guide
  • Neutral thread. Every bar is the same soft ink, for every nutrient and every amount. A thin tick marks the reference.
  • Said, not shouted. Past the reference, a sentence says “about 120% of the reference”. Never “over budget”, never “remaining”.
  • Honest about guesses. Estimates carry “~”, gaps in the data are named, and a portion can stay a range.
  • Yours to hide. One switch turns the numbers off if they feel like too much. The plate still lands.
The same lunch in the app: lentils 180 g and tomato 80 g, energy 240 kcal, protein 19 g and fibre 16 g, with nutrient bars against the reference intake.

A question, not a verdict.

After a plate, the coach may ask one thing, with a specific number to hand. Answer, ask it something back, or let it go. Dismissing costs nothing, and it never nags you to come back.

  • No streaks
  • No scores or grades
  • No badges
  • No countdowns
  • No red “over” states
  • No push to come back

Seven towels, side by side.

The week is read the same way as the day: which groups showed up, and which didn’t. Then the daily averages, plainly.

Veg and fruit showed up on six days this week, dairy on four. A sentence, not a score · demo week

Trust is the architecture.

There is no account and no server holding your meals. Your history is an encrypted database in your browser, opened by your passphrase. Here is everything that can leave the device, and when.

What leaves the device

  1. GET /api/v2/product/<barcode>?fields=…

    Only when you scan a barcode. To Open Food Facts, no cookie, no referrer.

  2. POST <your AI endpoint>

    Only if you turned an AI on, and only after you agree. A fixed set of fields; photos have their metadata stripped.

  3. GET /sw.js, /precache-manifest.json

    The app updating itself from your own server. It waits for you to tap Reload.

  4. telemetry, analytics, crash reports

    None. A build gate fails if one is ever imported.

The browser enforces this too: the content security policy blocks every other origin.

Defaults, drawn on off

  • Encrypted local storeOn
  • Ask before every photo uploadOn
  • Works offlineOn
  • AccountNone
  • TelemetryOff
  • AIOff until you choose

Your passphrase is the key. The store is SQLite encrypted with ChaCha20-Poly1305, its key derived from your passphrase on the device. A wrong passphrase cannot open it, locking drops the key, and erasing removes every byte.

Backups stay sealed. The backup file is sealed with AES-256-GCM under a backup passphrase (Argon2id). Restoring needs nothing but the file and you, on any instance, offline.

Every promise has a test. The README’s trust table names the test that proves each line, and the release fails without it.

The AI is optional. The app is complete without it.

Typing, barcodes, recents, saved meals and recipes all work with no AI at all. When you want photo recognition, descriptions or a conversation with the coach, pick where it runs.

  1. No AI

    Everything that doesn’t need a model. Nothing is sent anywhere.

    Default
  2. Your server’s AI

    Whoever runs your instance can connect a model once for the household. You still agree per photo.

    Optional
  3. Your own key

    Bring a key for any OpenAI-compatible endpoint. It lives in the encrypted store, bound to its host.

    Optional
  4. A local model

    Running a model on the device itself.

    Planned

Yours to run. One command.

The server is a hardened static container: read-only, non-root, no database, no outbound need. It holds none of your data. Put it on a home server or a small VPS, then open it on every phone in the house.

  1. Start it. One command builds and serves it. Add --tls for automatic HTTPS, or --traefik behind your own proxy.

  2. Open it on your phone. Over HTTPS, then Add to Home Screen or Install app. It works offline from then on.

  3. Choose a passphrase. It opens your encrypted store on this device. Each person, each device, their own.

On your server
# build and serve on http://localhost:8080
$ bin/ai-nutrition-coach up

# automatic HTTPS for your domain (Caddy)
$ export AI_NUTRITION_COACH_DOMAIN=plate.example.com
$ bin/ai-nutrition-coach up --tls

# or the published image
$ docker pull \
    ghcr.io/mmornati/ai-nutrition-coach
Read the self-host guide

Installs as an appA PWA with its own service worker; no app store.

Offline firstLog in a cellar or on a plane; nothing waits on a network.

English and FrenchFollows your device, or choose in Settings.

Accessible48px targets, screen-reader parity, reduced motion respected.

The short tour.

Drawn frame by frame in your browser, narrated, with subtitles. Shorter how-to episodes live in the docs.

All the how-to videos

Read the guide.