Flat vector motion graphics
Describe a clip.
Watch it get drawn.
Type a brief. A script model returns a scene — primitives with keyframed properties, not SVG and not a video — and your browser draws every frame from that description. No render farm, no queue, no upload.
- Drafted by
- DeepSeek V4.1 Flash, routed through OpenRouter with no fallbacks and no retention
- Cost
- A flat fifteen cents a clip, reviewed before anything runs
- Output
- A playable clip, a downloadable scene, and a link that carries the clip itself
The desk
Checking the lane…
Every figure in the trace above is read from the real response. Nothing there is an animation of work that did not happen.
A clip is about four kilobytes of geometry. That changes what it costs to make one.
One
The model writes a scene, not a picture
It returns rectangles, ellipses, polygons, paths and text with keyframed position, scale, rotation, opacity and colour in a fixed hundred-unit design space. Token count tracks the complexity of the scene, not the number of frames, so a five-second clip costs a fraction of a cent to draft instead of a dollar.
Two
This app draws every frame itself
The model never authors SVG. Constraining what it can return to a closed set of primitives is both a cost decision and a safety one: nothing it writes is ever parsed as markup or executed, and the only string that reaches a path parser is already limited to the path grammar.
Three
The clip is data, so the link is the asset
A whole scene fits in a URL fragment, which means a shared clip is re-rendered by the reader at whatever size their screen is, and keeps playing long after the sender has closed their laptop. Nothing is stored on a server to expire.