The basic idea is that everything on screen is a pure function of the current frame, so rendering frame N always produces the same result. That means you can seek anywhere in the timeline, and exports are frame accurate without needing a server.
Demos (each includes its own source): https://superhq-ai.github.io/webmotion/
The repo also includes an agent skill that lets coding agents create launch videos from a prompt. Quite a few of the demos were made that way.
Feedback is very welcome :)