HyperFrames (github.com/heygen-com/hyperframes) is an open-source framework, originated by HeyGen, that turns HTML, CSS, media, and seekable animations into deterministic MP4 videos. It renders with headless Chrome plus FFmpeg. You can use it locally via its CLI, drive it from AI coding agents through installable skills, or run it as the rendering core behind hosted authoring workflows. It's licensed Apache-2.0 with no per-render fees or commercial-use thresholds.

How is HyperFrames different from a normal video editor?

It's 'video as code.' Instead of a timeline GUI, you describe a video in HTML/CSS/JS with data-* timing attributes and 'clip' classes, then render it to MP4. Because the source is plain text, you can version videos in git, diff and review changes, template them, and generate hundreds of variants programmatically — none of which a traditional editor or render-queue SaaS makes easy. Rendering is deterministic, so the same source always produces the same frames.

How do AI agents use HyperFrames?

HyperFrames ships 21 installable 'skills' that agents load on demand. You run 'npx skills add heygen-com/hyperframes' and then describe the video; the '/hyperframes' router skill picks a creation workflow (product-launch video, faceless explainer, PR-to-video changelog, motion graphics, music-to-video, slideshow, and more) and teaches the agent the production loop: plan, write valid HTML, wire seekable animations, add media, lint, preview, and render. It works with Claude Code, Cursor, Gemini CLI, Codex, and other skill-supporting agents.

Is HyperFrames really free, and what's the catch?

Yes — it's open source under Apache 2.0 with no per-render fees or commercial-use thresholds. The main practical requirements are Node.js >= 22 and the headless-Chrome + FFmpeg rendering stack. It is NOT HeyGen's AI avatar/talking-head generation product; HyperFrames is the HTML-to-video rendering framework (the compositions behind HeyGen's own product launch videos). Think of it as an open engine for programmatic video, not a one-click avatar generator.

How does HyperFrames compare to Remotion?

Both render video with headless Chrome and FFmpeg, and HyperFrames is openly inspired by Remotion. The key difference is HyperFrames is HTML/CSS-first (rather than React-first) and explicitly designed to be agent-native, with a skills system that teaches AI coding agents the whole production workflow. It even ships a '/remotion-to-hyperframes' skill to port existing Remotion React compositions one-way into HyperFrames HTML.

HeyGen Open-Sources HyperFrames: Video as Code — Write HTML, Render MP4, Built for AI Agents

By Prahlad Menon Published 2026-09-20 3 min read

Every so often a big AI-video company open-sources the thing you’d expect them to lock behind a paywall. That’s what happened here: HeyGen has open-sourced HyperFrames — the “video as code” framework that renders MP4s from plain HTML — under a permissive Apache-2.0 license, with no per-render fees and no commercial-use thresholds.

The one-liner on the repo says it all: “Write HTML. Render video. Built for agents.”

What it actually is

HyperFrames turns HTML, CSS, media, and seekable animations into deterministic MP4 videos, rendering with headless Chrome and FFmpeg. You describe a composition in markup — using data-* timing attributes, class="clip", tracks, and sub-compositions — and it renders frame-accurate video. Because the render is deterministic, the same source always produces the same output.

That “video as code” framing is the whole point, and it unlocks things a timeline editor simply can’t:

Version videos in git — diff, review, and roll back changes like any other code.
Template and parameterize — swap variables to generate hundreds of variants in one run.
No editor, no render queue, no seat licenses — it’s a library and CLI you run yourself.

One important clarification: HyperFrames is not HeyGen’s AI avatar/talking-head product. It’s the rendering engine — in fact, it’s the framework behind HeyGen’s own product-launch videos — now handed to the community.

Built for agents, not just developers

The most forward-looking part is the skills system. HyperFrames ships 21 installable skills that AI coding agents load on demand:

npx skills add heygen-com/hyperframes

Then you just describe what you want — “Using /hyperframes, create a 10-second product intro with a fade-in title, a background video, and subtle music.” A /hyperframes router skill reads your intent and picks the right creation workflow, of which there are many:

/product-launch-video — market a website/product from a URL, brief, or script.
/faceless-explainer — explain a concept with LLM-invented typography and data-viz.
/pr-to-video — turn a GitHub pull request into a changelog/feature-reveal video (read via the gh CLI).
/motion-graphics, /music-to-video (beat-synced), /talking-head-recut, /embedded-captions, /slideshow, and a /general-video catch-all.

The skills teach the agent the full production loop — plan, write valid HTML, wire seekable animations, add media, lint, preview, render — and they work with Claude Code, Cursor, Gemini CLI, Codex, and other agents that support skills.

It’s the Remotion lineage, HTML-first

If this sounds familiar, it should: HyperFrames is openly inspired by Remotion, and both render video via headless Chrome + FFmpeg. The distinction is philosophy. Remotion is React-first; HyperFrames is HTML/CSS-first and agent-native, designed so an LLM can author compositions directly. There’s even a /remotion-to-hyperframes skill for one-way porting of existing Remotion React compositions.

Is the hype warranted?

Mostly, yes — with the usual grounding. The genuinely exciting part isn’t “make one video faster”; it’s programmatic video at scale: content automation, batch generation, data-driven videos, and text-to-renderable-video pipelines that agents can run end to end. For anyone doing marketing at volume — 200 localized ad variants, a changelog video per release, a data video per customer — that’s a real unlock, and having it open source under Apache-2.0 removes the per-render economics that usually kill those ideas.

The realistic caveats: it’s a developer framework (Node.js ≥ 22, a Chrome+FFmpeg render stack), not a drag-and-drop app, so the “no editor” pitch cuts both ways — you trade the GUI for code. And “video” here means HTML/CSS/JS motion design and composited media, not generative AI footage. But as infrastructure for the emerging world of agent-driven content production, HyperFrames is one of the most consequential open-source releases in the space this year.

Repo: github.com/heygen-com/hyperframes · Site: hyperframes.heygen.com