Most cooking apps want three things from you before they’ll suggest dinner: an account, a connection, and a slice of your attention to sell to someone else. PantryAtlas wants none of them. It runs entirely on a Raspberry Pi 5 sitting on your counter, it never phones home, and the only question it asks is the one that actually matters at 6pm on a Tuesday: what do you already have?

PantryAtlas is live today at pantryatlas.org, open source under Apache 2.0 at github.com/PantryAtlas/pantryatlas, and tagged v0.2.0.

The whole idea

Tell it what’s in your pantry. Get back recipes you can actually make right now, ranked best-first. That’s the entire loop — pantry-in → ranked-recipes-out — and it happens on a computer that costs about $80, with no internet connection required after setup.

Under the hood it’s matching your ingredients against 49,965 recipes, scoring each one on how much of it you can already cover, what’s about to expire, how few substitutions you’d need, and whether it fits the kind of cooking you do. The top of the list is the thing worth making tonight.

You talk to it from any phone or laptop on the same Wi-Fi by opening pantryatlas.local in a browser. It installs as an app (it’s a PWA), works with the network unplugged, and keeps your pantry on the device — not on anyone’s server.

Two ways to add what you have

• Type it. Start typing “tomatoe” and it resolves to tomato — exact match first, then fuzzy spelling, then meaning. It speaks 100+ languages, so tomate and tomatillo land where they should.
• Photograph it. Point your phone at the shelf and PantryAtlas reads the ingredients straight off the image, using an on-device vision model. No upload, no third-party API.

Who it’s actually for

I built the ranking with community kitchens in mind — food banks, shelters, shared pantries — places where “what can we cook with the donations that came in today, before the produce turns?” is a real daily question, and where sending the inventory to a cloud service is a non-starter for cost and privacy reasons. The expiration term in the ranking isn’t decoration; it’s there so the thing that needs using up floats to the top.

But it’s just as happy being the small box in your home kitchen that turns “half an onion, a can of beans, and some rice” into three things worth eating.

Getting it running

Two paths, depending on whether you like flashing SD cards:

The fast way — flash an image. Download the prebuilt SD card image (~1.9 GB), write it to a microSD card with Raspberry Pi Imager, put it in your Pi 5, and power on. About eight minutes of writing, one minute of booting, then open pantryatlas.local. Steps are in the Download & flash guide.

The hacker way — bootstrap a fresh Pi. If you already have Raspberry Pi OS running:

git clone https://github.com/PantryAtlas/pantryatlas.git ~/pantryatlas
bash ~/pantryatlas/ops/pi-bootstrap.sh
# ~30 min later: "BOOTSTRAP COMPLETE"

Either way, the heavy lifting is handled for you: the prebuilt recipe database (~227 MB) downloads automatically on first boot and checks itself against a published checksum, so you skip the multi-hour job of building it locally. (More on how I ship that file for essentially nothing in a companion post.)

What “local-first” actually buys you

It’s easy to say “no cloud” like it’s a feature checkbox. Here’s what it concretely means for PantryAtlas:

• Your pantry is yours. The list of what’s in your kitchen lives in the browser on your device and on the Pi. It is never transmitted to me or anyone else, because there’s nowhere for it to go.
• It works when the internet doesn’t. A community kitchen with flaky Wi-Fi, a cabin, a basement — once it’s set up, the recipe brain is right there.
• No subscription, no rug-pull. It’s an open-source package you run yourself. There’s no server I can shut down and no plan I can raise the price on.

Where it’s headed

v0.2.0 is the first release with the full app on top — earlier work built the primitives (embeddings, storage, the Gemma wrapper). Next up: wiring live on-device vision end to end, smarter filtered search, and a multilingual recipe vocabulary so the 100-language embeddings have 100-language recipes to match against.

If you want the engineering story — how you rank 50,000 recipes on a CPU that can only embed about 16 short phrases a second — that’s the next post. If you want the docs, they’re at docs.pantryatlas.org. And if you’ve got a Pi 5 in a drawer, you know what to do with it now.