The lesson from Brandblast
I left Meta to build Brandblast, an AI content product for brands. Within six months of leaving it was doing $17k MRR, and it went on to serve more than 5,000 brands.
Watching that volume of AI-generated content ship every day taught me the thing Clipkit is built on: the models were not the bottleneck. The tools were.
Every tool failed one of two ways
I was building on the bleeding edge of the existing editors and video frameworks, and they split cleanly into two camps.
On one side were great editors sitting on poor source formats: proprietary markup that represented the media but that no AI could reliably read or write. On the other side were templating systems with beautifully simple markup but weak editors or none at all, and output quality that capped what you could express.
Great surface, illegible substrate. Legible substrate, weak surface. Nothing gave you both.
The bet
So I decided to build the complete system: a simple JSON protocol that maps directly onto WebGPU and WebGL2.
Any AI can compose video by writing a document. Anyone who wants to build a video editor can build on top of it: their UI just edits the source JSON, and the renderer takes care of the pixels. The document stays legible to machines, the surface stays great for humans, and neither one compromises the other.
Later I added a Rust wgpu renderer that powers our enterprise deployments and cloud rendering, so the same protocol renders identically from a browser tab to a server GPU.
Where it's going
The goal of Clipkit is simple to state: anyone, human or agent, can compose video, with no infrastructure getting in the way. A holistic system with no limitations of tech stack or environment.
The protocol is open, the editor is open source and embeddable, the no-login editor runs in your browser right now, and the architecture page shows how the whole stack fits together.
If you hit a wall anywhere, email me: support@clipkit.dev. I read everything.