Mise

Inspiration

A simple tomato pasta can be the start of learning to cook, not just a recipe to finish. Mise explores how to make the next action manageable and add a cooking technique when the learner is ready.

What it does

Four cooking lines—pasta, rice, potato and one-pan oven—each contain ten recipe versions. Servings and dietary preferences change ingredients and directions together. People choose guidance, use one-step mode, pause and return, and keep approved preferences in their account.

Why WebMCP

The assistant acts on the same cooking workspace the person sees. Eleven page tools read current state, compose a session around multiple constraints, propose smaller actions, and request human confirmation. The assistant cannot award completed cooking through a tool call alone. Suggested memory stays pending until the person approves it.

How it was built

React, Vinext, Cloudflare Workers and D1, with imperative document.modelContext.registerTool registration. Runtime schemas reject malformed inputs. Context tokens prevent stale recipe changes; per-account storage and revision checks protect saved progress.

Challenges and lessons

Testing found that paused sessions could be changed invisibly and that pending completion could outlive a servings change. Both were fixed and retested. A final beginner-rice correction made adding water, boiling and switching heat off explicit. Shared state and learner control matter more than simply awarding levels.

Validation and boundaries

36 automated tests, lint, typecheck and build passed. Actual browser checks covered shared state, persistence, pending approvals, invalid tools, mobile layout and keyboard navigation. Six cooking personas are simulations, not real-user research. Recipes are not professionally cook-tested or allergy-certified. This is browser WebMCP while Mise is open, not a remote connector in every GPT or Claude chat. Separate-chat briefs are manual snapshots, not automatic account sync.

Release status

Cooking-only version 9 matches source 9a4a6d7dc64edfe3e8bc12575d63b34ac871cae8. The app and matching MIT-licensed source are public. Anonymous account requests are denied; each person keeps their own saved record. The submitter confirms the food images were AI-generated in this project. The demo uses synthetic English narration and edited captures of actual live app interactions in temporary mode.

Built With

• react