JSX-as-timeline is a smarter format for agent-driven video editing than any GUI

The core bet here is that JSX is a better document format for video than a timeline GUI. A project is a folder of SolidJS components: a <stage> holds <scene> elements, clips and text carry timing as props, and generative assets — images, video, voiceover — are declared inline with generate.* calls. Save the file and the editor remounts the tree. Drag a clip on the canvas and the change writes back to the JSX source. The document and the UI are the same thing, kept in sync through an ECS runtime that has no DOM dependency.

What makes this interesting for agent workflows specifically is the dapi CLI, which is the real interface an LLM actually touches. The command set is designed around the problem that an agent cannot watch footage. dapi media grab decodes frames to PNGs at specified timestamps. dapi media filmstrip produces a grid of frames. dapi media waveform renders an audio waveform with silence flagged. dapi media transcribe returns a word-level timed transcript. dapi media listen lets you ask a multimodal model a question about the audio directly. dapi capture renders what the timeline would actually produce at given positions — so an agent can verify its own output without exporting. Every command returns JSON or JSON Lines on stdout, errors go to stderr with exit code 1, and the whole thing is built to be piped.

The architecture separates concerns cleanly. The headless runtime (@diffusionstudio/runtime) has no DOM and no Solid dependency — it is a koota ECS world with media decoding and capture. The reconciler evaluates a compiled project bundle and maps its element tree onto runtime entities using Solid's universal renderer. The encoder handles offline export. This means the runtime can run without the UI, which is what makes CLI-driven workflows viable.

The prompt examples in the README are instructive about the intended use. One asks for a recreation of a 3Blue1Brown animation using the exact collision mathematics from a specific academic paper. Another asks for a 30-second vertical clip pulled from a YouTube URL. These are not toy demos — they describe a workflow where an agent researches source material, generates assets, assembles a composition in JSX, and the human reviews the result in the editor before exporting.

The desktop download is currently macOS Apple Silicon only, though the web app at app.diffusion.studio is available as an alternative. The project is MPL-2.0, with brand assets carved out separately, and is a Y Combinator F24 company.