HyperFrames

Programmatic video has long forced developers to choose between two awkward options: learn a proprietary timeline format and editor, or stitch frames together by hand with brittle scripts. HyperFrames, an Apache-2.0 project from HeyGen, takes a different starting point. It is a framework that turns HTML, CSS, JavaScript, and seekable animations into deterministic MP4 video. You author a composition the same way you build a web page, mark up timing with data attributes, and the renderer produces a frame-accurate clip. The explicit design goal is to make video something an AI agent can write, since any model that can produce HTML can produce a HyperFrames composition without learning a new domain-specific language. That framing, video as code rather than video as a tool, is what sets it apart from conventional editors.

The pipeline is conceptually simple, which is part of its appeal. HyperFrames parses an HTML composition where data attributes define tracks, frame ranges, and timing. It then drives headless Chrome through Puppeteer, seeking each animation to a precise timestamp and capturing the frame at that exact moment. FFmpeg encodes the captured frames and mixes any audio tracks into the final MP4. Because every frame is rendered by seeking to a fixed position rather than recording playback in real time, the output is deterministic: the same input yields the same frames every time, which makes visual regression testing in CI feasible. Animation support is adapter-based and covers GSAP, Lottie, Three.js, Anime.js, CSS animations, and the Web Animations API. The one hard requirement is that animations must be seekable.

The clearest audience is teams building automated, data-driven video pipelines, especially ones where an agent generates the composition. HeyGen positions it for AI coding tools such as Claude Code, Cursor, and Gemini CLI via installed skills, but it works equally well as a plain CLI for developers rendering locally. Concrete use cases from the project include templated product demos and feature announcements, pull-request walkthroughs with animated code diffs and narration, animated charts and data visualizations, and social clips with captions. The common thread is repeatable video produced from structured input at volume, not one-off creative pieces. If you already think in HTML and CSS and want to generate hundreds of personalized or templated clips programmatically, the authoring model will feel natural rather than foreign.

The honest caveats are real. Running headless Chrome plus FFmpeg is meaningful infrastructure weight for self-hosters: you need Node.js 22 or newer, a working FFmpeg install, and enough compute to render frames at acceptable speed. The repository also relies on Git LFS for regression baselines, which the project reports as roughly 240 MB to clone. Maturity is another concern. At v0.x with a very high release cadence, the API is explicitly still considered unstable, so pinning versions and budgeting for breaking changes is prudent. The roadmap is steered by HeyGen's commercial priorities, which is reassuring for longevity but means feature direction may not always match an independent adopter's needs. Finally, this is a code and agent interface: there is no WYSIWYG canvas, so non-developers cannot author with it directly.

Against its commercial relatives, the trade-off is clear. Managed services like HeyGen's own platform or Remotion Cloud remove the infrastructure burden and add hosting, scaling, and support, at the cost of per-render fees and less control. HyperFrames gives you the rendering engine itself under Apache-2.0 with no per-render cost, which is compelling if you can run the stack and want full ownership of the pipeline. Adopt it when you are generating templated, data-driven, or agent-written video at scale and are comfortable operating headless Chrome and FFmpeg in CI or on Lambda. Treat it as too early if you need a visual editor, non-developer authoring, or a frozen API today. For its intended niche, agentic video as code, it is one of the more credible open options, with the usual v0.x volatility to plan around.