autoPlan

Every household deals with it: what to eat next week, who's shopping, what's already in the fridge. It's usually an hour of back-and-forth that rarely produces a plan lasting past Monday - low stakes, high friction, and a good fit for automation.

Instead of asking the household to plan, autoPlan asks them to react. An AI agent drafts a preference-aware week from the recipe library and drops it into the group chat everyone already uses. Approve keeps it, swap rerolls a night, veto removes a dish for good. The plan settles in a few taps.

The planning agent runs on a weekly schedule and talks to the household through a Telegram bot that polls for votes. Supabase is the single source of truth - recipes, preferences, cook history, and proposals all live there, and every consumer reads it directly. Once the plan locks, a cart-builder agent logs into Instacart with Playwright and builds carts split across the household's regular stores, then hands back per-store totals for a final review. The AI never places an order - it only gets the carts ready.

There are no servers to maintain. The agent runs on a schedule, the bot polls instead of holding a webhook, and the data layer is a managed Postgres on a free tier.

The loop runs unattended every week. The recipe library, preferences, and cook history are its memory, so the drafts get better and the vetoes get rarer. The negotiation didn't get faster - it went away.