JellyFlow/Mcdonalds_in_RokidGlasses

This is a contest entry that lets a Rokid Glasses wearer order McDonald's by speaking, while keeping门店 selection, pricing and the final submit step under explicit user control. Try it if you own the hardware and want a concrete example of a smart-glasses MCP workflow; skip it if you need a finished, fully validated ordering product.

The problem

Ordering food on a phone is awkward when your hands are busy or the phone is in a pocket. A glasses wearer wants to say something like "two Big Macs and a fries, pickup at the store," then check the store, item details and total on a small 480×352 display. The project splits that into: speech captures the intent, the store is selected and confirmed, the menu is searched through the real MCP service, and the cart is priced from returned values. The result is a short, glanceable flow where the user still decides what actually gets submitted.

What's different

The design is deliberately conservative about automation. Each menu page holds at most five items with numbered labels, and local batch switching avoids repeat MCP requests. Ambiguous or duplicate dish names are handed back to the user rather than auto-added, and when a price cannot be matched reliably the UI shows "price to be verified." Deletion, quantity changes and a single-level undo of the last removal cover cart mistakes. The author frames this as preparing the flow by voice, then requiring a key press for new addresses, coupons, order creation and cancellation.

The trade-off is that the device talks directly to McDonald's MCP in the real mode, so the desktop does not need to stay on, but every meaningful state change still needs eyes on a tiny screen and a physical confirm. That is a sensible boundary for a prototype, though it means the voice interaction is an accelerator rather than a hands-free ordering system.

Who should try it

Best fit: a Rokid Glasses owner and McDonald's MCP developer who wants a working reference for AIUI pages, device-side MCP calls and voice-plus-key interaction. The mock path is also useful for someone evaluating the interaction design without the hardware. Poor fit: anyone expecting to install an app and order dinner without their own MCP token, or someone whose primary goal is a general smart-glasses assistant rather than this single branded flow. The project is explicitly an independent contest entry, not an official McDonald's or Rokid product.

Requirements and limitations

Running the real flow needs Node.js 22.13+, your own McDonald's MCP token from the MCP platform, a configured local environment per the development guide, and an authorized ADB-connected device for installation. The isolated mock route needs the aix CLI and, per the README, a macOS Chrome path by default; other systems must adjust the script. The repo does not ship a personal AIX package or credentials, and a locally built AIX contains your device credentials and must not be redistributed.

Validation is partial by the author's own account: the verification record covers some real-device queries, voice selection and cart operations, but not a full new-account install-to-order acceptance. The README states that real coupon claiming, address addition, ordering and cancellation are not executed in the public cleanup, and payment stays in the phone app. Menu, prices and coupon availability depend on the store, account and network, so behavior will vary. Whether the "double-click order status to return to system desktop" path is accepted is also flagged as unverified.

HotRepo's take

HotRepo first tracked this repository on 2026-10-24 and observed it only once, on the 2026-10-25 NEW-REPOSITORY board (新项目榜, not the overall Trending board) with 114 total stars at #12 in the TypeScript filter. That single dated snapshot says nothing about momentum, and persistence cannot yet be judged. Before investing time, verify that your MCP token account and glasses firmware support the documented AIUI/MCP flow and that your device model is covered by the verification record.