PantryOS

Dashboard and AI activity
Saved recipe
Kitchen inventory
Recipe matching
Grocery planning

Inspiration

I do a ton of home cooking and managing the food in the fridge, can be a real hassle from what Ingredients I have to if they are expired or not. Inventory, recipes, and grocery lists usually live in separate places, leaving me having to compare dates, quantities, and missing ingredients manually.

PantryOS explores a different model: Where I already implement AI for planning meals and recipes and managing leftover ingredients it instead homogenizes the information so an agent can access all of it at once.

What it does

PantryOS tracks ingredients across the fridge, freezer, and pantry, including quantities and expiry dates. It prioritizes food that should be used soon, finds suitable recipes, scales serving quantities, converts units, and can add missing ingredients to the grocery list.

Instead of navigating several screens or finding recipes or copying information down and adding it to boxes of text manually the AI agent can automate this for you navigate fill out grab info from and expedite all this busy work so you can get back to your daily life. Here's an example of a use case

Find a recipe for four people that uses the ingredients going bad, then add anything else I need to my shopping list.

PantryOS turns that request into actual management actions keeping the interface and stored kitchen state up to date as you work or cook.

Why WebMCP

PantryOS is not a chatbot taped together alongside an application to manage your kitchen. Its features are exposed using structured WebMCP tools.

The agent can inspect inventory, evaluate expiry dates, save and retrieve recipes, update quantities, and plan groceries without you having to feed it information about your kitchen and the recipes you have every time.

What you end up with is a superior user experience, people can talk to agent in everyday language while PantryOS handles the multi-step coordination between inventory, recipes, serving sizes, and groceries. This also allows for further and more complex syncing of your information later on, the easy of use of the MCP allows for later hosted version of this demo to enable access by your AI agent wherever you are on the go.

Human-agent collaboration

The agent can prep and perform useful kitchen actions, but the person remains in control. Destructive actions like deleting recipes or throwing out food require confirmation to make sure the Agent doesn't run haywire and delete your grandmother long lost recipe. The activity panel records recent agent assistance, making it clear what is happening and what changed.

How we built it

PantryOS is a client-only React and TypeScript application deployed using Cloudflare Workers Static Assets. Zustand persists inventory, recipes, and groceries in browser localStorage.

Its WebMCP tools register through document.modelContext. Every tool input is validated with Zod and described using a matching closed JSON Schema. Tool callbacks read current application state when they execute and call the same domain actions used by the interface.

Automated tests cover tool contracts, persistence, malformed inputs, quantity conversion, grocery deduplication, cancellation, stale state, and confirmation races.

What we learned

The best agent tools don't feel like a background script or an entirely separate app. They work best when they can work with what you already have and you can see clearly their application.

WebMCP made this feel like a more native interaction that just a automations or AI agent that is just added to a kitchen management app. Its quick simple to use and setup easy to for the AI to work with and access the correct information and allows me to use my already exiting workflow and models to do my tasks.

What's next

Future versions could add opt-in household collaboration and import integrations. I would like to have it persist across different browsers so you can access the info on the go. The submitted version intentionally remains local-first, client-only, and account-free so the WebMCP interaction is focused, private by default, and easy for judges to reproduce. But an increasing in scope and features is quite easy as there are tons of features I could have added that would be useful to me reminders, logs, side notes for recipes and so on that normal people use and interact with the kitchen everyday.