Framefields, an open-source project on GitHub, lets an AI coding agent build a video by writing code instead of dragging layers around an editor. The agent describes the scene in TypeScript, renders sample frames to check its work, then exports an MP4.
Video has been hard for agents because editing tools are built for eyes and a mouse. A code interface with frame previews gives a model something it can read, change and verify.
The README says it draws on the graphics chip through WebGPU (a browser standard that lets code use the graphics hardware directly) instead of driving a hidden browser. It describes its scope by analogy to Photoshop, After Effects and Blender. Those are the project’s own descriptions of scope, not claims of parity.
The README opens with a beta warning: the project is under active development, interfaces can shift between versions, and some features may be unfinished. It names no users and no adoption figures, and its speed and memory comparisons with rival tools are its own, unevaluated. Treat it as a prototype to test, not a pipeline to depend on.
Source: the Framefields project README on GitHub (gatewai-dev/framefields), which carries no author and no publication date.
