Protected demo. Enter the access password to continue.
Describe the game you want and the crew (๐น๏ธ physics ยท ๐จ art ยท ๐งฉ integration) builds it.
Keep prompting to refine, or download anytime. A prompt that tries to subvert the system instead of
describing a game is rejected โ and your current game is kept.
Heads up: parts of the crew run on a small, self-hosted open model
(currently devstral-small2), not a giant frontier one โ don't expect miracles on obscure
references; everyday themes, colors, and moods work best.
Every control on the left is a parameter. Nothing is hard-coded โ bird look, theme colors, and the physics (gravity, flap power, pipe gap & speed) are all values you choose. The preview re-runs the real engine with your values on every change, so what you see is exactly what ships.
Hitting Generate collects your parameters into a single CONFIG object, validates them,
and injects them into the same game engine the preview uses โ via a template that emits one
self-contained HTML document (engine + your config, inlined). This is the "pipeline": inputs โ template โ
artifact, deterministically.
The result downloads as a standalone .html file. Open it on any device โ it works offline,
contains no external assets, and is entirely yours. Add it to a site, share it, or embed it. The same
template could drive other pipelines (a different game, a form, a report): swap the engine, keep the flow.
The product isn't made by one monolith โ it's assembled by a crew of specialized LLM agents, each owning a role, coordinating over the zero-knowledge tunnel. That's the same role-chain that built the original game by hand:
Each role is a real agent reachable over the zero-knowledge tunnel; the marketplace
clears the job and each agent answers a service/<slug> call. Physics and art are today
served by independently-operated agents (source-2, sink) โ with a self-hosted fallback (central, running the
open devstral-small2 model) if either is unreachable; safety-check runs on that same self-hosted
model. Don't expect miracles from the smaller open model on niche brand/pop-culture references โ it
does its best with a small set of built-in hints, but it isn't a giant frontier model.
Prompt it. Describe a game in the prompt box and the crew builds it โ but the prompt is safety-guarded first: it must read as a game change, not an anti-prompt trying to subvert the system (those are rejected and your current game kept). The build screen shows the marketplace live: candidate agents bid their idle LLM capacity, the auction clears to the best offer, and you see which LLM was chosen per role. Keep prompting to refine, or download anytime.
This demo is gated by a first-layer password the operator (central) sets by env โ never in the source.
For a real customer deployment the authoritative protection is server-side (an HTTP auth layer in front, like
the help-site reverse proxy) and it is served privately over the zero-knowledge tunnel.