Grocerly

The problem

Recipe apps don't know what's in your pantry, and grocery-list apps don't know what you're cooking. The result is the same list re-typed every week and a second jar of something you already had.

Grocerly connects the two: track the pantry once, and both the recipe library and the grocery list work from it.

How it works

Ingredients live in categorized pantry sections, each marked have, need, or don't have. Recipes come in from a link, pasted text, a file, AI generation, or shared straight in from other apps, and the library filters down to what you can cook now — or are only an ingredient or two away from.

The grocery list is the gap between a recipe and the pantry. An AI chef, which you can name yourself, knows your pantry, preferences, and allergies, reads photos, and comes along into a Cooking Mode that keeps the screen awake and runs timers. Macros scale with servings.

Households share one pantry and grocery list across a family, and friends can send each other recipes, with push notifications on iOS and Android.

Engineering decisions

1. One codebase, three storefronts

   The app is a React 19 + Vite web app, wrapped by Capacitor for iOS and Android. One build is copied into both native shells, so a fix ships everywhere from a single change.

2. Local-first where it can be, shared where it has to be

   Your own data lives on the device first and syncs to Firestore in the background, so the app opens instantly and works offline. Household data is shared, so it lives in Firestore with an on-device cache and merges field by field — two people editing the same list don't overwrite each other.

3. AI behind a server boundary

   Every model call goes through serverless functions that hold the API key. Subscription status, daily quotas, and AI consent are all enforced on the server; a quota slot is reserved before a paid call and refunded if the call fails.

4. Recipe import that can't be turned against the server

   Importing from a link means the server fetches addresses a user chose. That fetch is HTTPS-only, blocks private network ranges, re-checks every DNS answer and redirect, and pins the connection to the address it vetted, so DNS can't be swapped mid-request.

5. Nutrition from USDA data, with AI only filling gaps

   Macros come from USDA FoodData Central; the model's job is matching an ingredient to the right USDA food. The rare line USDA can't resolve gets an AI estimate, and the app labels it as one.

6. Households are created server-side

   Clients can't create a household directly. A server function checks the subscription and creates it with admin privileges, so the rules on who can create one can't be bypassed from the device.

7. Translation at render time

   The interface ships in English, Spanish, and Russian. Categories, cuisines, and meal types are stored as canonical keys and translated only when displayed, so the same household reads correctly in every language.

Stack

Frontend

React 19, Vite

Native

Capacitor (iOS + Android)

Backend

Vercel serverless functions

Auth & data

Firebase Auth, Firestore, Storage, Cloud Messaging

AI

Anthropic Claude

Nutrition

USDA FoodData Central

Subscriptions

RevenueCat (App Store + Google Play billing)

i18n

i18next

Testing

Vitest, Testing Library, Stryker, Firestore rules emulator