LevelDez

A browser 3D interior level designer that furnishes itself: one sentence generates corridors, halls, and richly decorated rooms, and the result exports as a JSON layout plus a multi-mesh GLB your engine can load.

Not per the captured README: it names Microsoft's Decision-1 as the small yes/no model that presses the buttons and never mentions a Jev API. It is cataloged from the Made with Jev gallery, where it appears as a System One-style build, so treat it as an adjacent decision-model project rather than a Jev client.

With Node 18+: git clone the LevelDez repo, cd LevelDez, copy .env.example to .env (keys are optional), and run npm start for http://localhost:5173. No npm install is needed; three.js loads from a CDN, so the first load needs an internet connection.

The README's decision card says Decision-1 gives calibrated yes/no probabilities, batched in one cheap call, and shines at taste-style room judgments while deterministic code handles geometry. A full autopilot run costs roughly a tenth of a cent.

One sentence can produce a whole building: corridors, side wings, big halls, and rooms with their own wallpaper or stone, floors, wainscot, trim, and coloured mood light, composed from about 600 hand-built assets. Exports are a JSON layout and a multi-mesh GLB.

The captured README was cut off mid quick-start, so the full walkthrough is incomplete; no accuracy figures or benchmarks are published, the author shares it as a personal project built in a few hours, and the first load needs internet for the CDN-hosted three.js.

It is yeme-oss's project, published under the Unlicense per GitHub metadata; the assets were made in Blender with BlenderMCP, three.js r160 is used, and npm dependencies are zero — one small Node server and static files.