Inspiration

Planning dinner and recording that dinner was cooked are different actions. An assistant should not subtract supplies just because someone asks what they could make. It should also share ingredients across the whole plan instead of promising the same tomato twice.

What it does

Pantry Before Shopping is a local MCP server with a guided browser demonstration. Four tools inspect recorded lots, preview dinners, confirm cooking and add supplies. A preview allocates lots across every selected dinner and totals missing portions. Date labels prioritize lots; they do not establish food safety. A plan with shopping gaps cannot be confirmed.

Confirmation consumes supplies and writes a receipt. A changed pantry revision invalidates earlier previews. Repeating confirmation cannot consume the same plan twice. State and receipts survive restarts.

Demonstration

The 51-second demo starts with two synthetic dinners for two and the exact combined gap: three tomato portions and one pasta portion. Enlarged app views show shared allocation, a preview that leaves stock unchanged, and an explicit cooking confirmation with a recorded receipt after missing supplies are added. It combines explanatory graphics, edited screenshots of real MCP-backed app states, and locally synthesized narration. Portions are household-defined recipe units. The guided deterministic demonstration is not a live Alexa conversation or a shopping order.

How it was built

Intended primary track: Alexa+, self-hosted MCP route. The official MCP SDK implements Streamable HTTP and negotiates protocol 2025-11-25. The browser initializes MCP and calls the actual tools directly. It is a guided deterministic interface, not a live Alexa connection, voice recognizer or language-model conversation. Another compatible MCP client can use the server. Node.js, Express, Zod, HTML, CSS and JavaScript support the implementation. No paid API, runtime model, shopping order or cloud account is used.

Challenges and validation

The key design challenge was separating a proposal from an inventory-changing action, while preserving an exact revision across multiple previews. Six automated domain/integration tests and nine browser checks passed, including a fresh checkout. Tests cover combined shortages, lot ordering, invalid inputs, stale previews, replay rejection, persistence, actual SDK client calls and Host/Origin checks. Browser checks exercise actual HTTP tools and a 390-pixel layout.

Limits and next steps

Initial pantry data is synthetic. Four curated recipe templates use household-defined portion units. There is no nutrition analysis, food-safety determination, automatic purchasing, cross-device synchronization, authentication for public hosting or measured household savings. This is a single-process local tool; compatible assistants must request explicit user approval before mutation tools. Independent household testing and a trusted multi-user approval system are future work.

Implementation, testing and these materials used AI assistance. No novelty, Alexa certification, contest acceptance or award probability is claimed.

Built With

Share this project:

Updates

Submission history