Write HTML, Get an MP4: HyperFrames Turns Video Into Something an Agent Can Build

Write HTML, get an MP4

HeyGen open-sourced HyperFrames this month, and the tagline is almost aggressively plain: write HTML, render video, built for agents. There is no diffusion model inside. Nothing is dreamed up. You write a web page, the page is loaded in headless Chrome, every frame is captured from a timestamp you asked for, and FFmpeg stitches the frames into a file.

That sounds like a step backwards in a year when video models can produce a convincing street scene from a sentence. It is actually the opposite. HyperFrames solves the half of video production that generative models are worst at: the part that has to come out the same way twice.

The project is Apache-2.0, sat at roughly 63,900 stars by mid-October after adding more than a thousand in a day, and pulled about 811,000 npm downloads in the week to October 12. It is, by a wide margin, the most adopted thing in the "agent makes video" category, and it got there without generating a single pixel of original imagery.

What deterministic actually buys you

A video diffusion model is a sampling process. Ask for the same prompt twice and you get two different videos, which is fine when you are exploring and terrible when you are shipping.

HyperFrames does the opposite. Each composition is an index.html. Clips are ordinary elements carrying data attributes for start time, duration, and track index. Animation comes from GSAP timelines, CSS keyframes, Lottie, Three.js, Anime.js, or the Web Animations API. The renderer loads the page, jumps to each frame by seeking rather than playing in real time, captures it, and hands the sequence to FFmpeg. Same input, same frames, every time.

That property is boring until you put an agent in the loop. An agent that changes one line of copy needs to know the rest of the video did not move. With a generative model, re-rendering after a text fix means re-rolling the whole clip and hoping the timing survives. With a deterministic renderer, the fix is a fix.

There is a second-order benefit that matters more for teams than for individuals: a video becomes a diffable artifact. You can see what changed between version three and version four, because it is HTML.

The skills are the actual product

The renderer is the easy part. The harder part is teaching a coding agent to write a composition that does not look like a web page with a video in it.

HyperFrames ships 21 agent skills, with /hyperframes as a router that reads the request and picks a workflow. The catalogue is shaped like real jobs rather than like features: a product launch video from a URL or a brief, a faceless explainer from text, a music-to-video pass that syncs cuts to beats, embedded captions, a talking-head recut, and short motion-graphics hits for logo stings and stat cards. The skills work with Claude Code, Codex, Cursor, and Gemini CLI.

The cleverest piece is frame.md. Plenty of brands now maintain a design.md, colors, type scale, spacing rules, that an agent can read. The problem is that those rules are written for screens. Applied to video, an agent will happily produce 14-pixel body text and a navigation bar, because that is what the rules say. frame.md rewrites the same brand system at video scale: type that reads at arm's length, compositions that survive a 16:9 crop, motion that has somewhere to go. The gallery ships a dozen starting systems, including a vintage sports-magazine look built around one saturated red.

This is the kind of unglamorous detail that decides whether agent-made video looks professional. A model that generates a beautiful frame does not know your brand has a specific red. A skill file does.

The Remotion comparison, stated honestly

HyperFrames is openly inspired by Remotion, the React-based video framework. The team used Remotion in production, kept attribution comments in the source for the patterns it pioneered, and made one different bet: Remotion asks you to write React components, HyperFrames asks you to write HTML.

The trade is clear. Remotion gives you a real programming model and a build step. HyperFrames gives you no build step at all, an index.html plays as-is, and leans on the fact that language models are extremely good at HTML and CSS, and merely good at React. For an agent-first tool, that is the right bet, because the failure mode you are optimizing against is an agent that writes plausible-looking JSX that does not compile.

Requirements are modest: Node 22 or newer, FFmpeg, Chrome. Each render worker uses around 256 MB of RAM. No GPU is required. Renders can run locally, on AWS Lambda, or through HeyGen's hosted service, which is where the money is: local rendering costs nothing beyond your own machine, hosted renders are billed per credit, and 4K is billed at 1.5×.

Version 0.8.145, shipped October 10, added GPU export inside HyperFrames Studio and fixed a bug that left long exports stuck at 90%, a small release note that tells you a lot about where the project actually is. This is not a research demo. It is a tool with a bug tracker and users who notice when a progress bar lies.

The honest limits

HyperFrames cannot invent footage. If you need a crowd scene, a waterfall, or a person's face doing something specific, this is the wrong tool and you should use a video model. It renders what you describe in code: typography, charts, UI walkthroughs, lower thirds, kinetic type, 3D objects you already have, video you already shot.

The output is also exactly as good as the HTML an agent writes for it. A weak agent produces a slideshow with fades. The skills raise the floor considerably, but they do not replace taste, and they will not save a prompt that never said what the video was for.

And the Studio desktop app now asks you to sign in with a HeyGen account, which is a reasonable business decision and also a reminder that the free part is the engine, not the whole product.

Why this matters more than another video model

The last year of AI video has been a race on the same axis: longer clips, better physics, more consistent characters. All of that is real progress, and all of it is about generation.

But a large share of professional video is not generation. It is a chart that has to animate correctly, a lower third that has to appear at 0:03, a title that has to be legible on a phone. Those jobs were never waiting for a better model. They were waiting for a way to hand the work to something that does not get bored and does not drift.

HyperFrames is that handoff, and its strongest claim is not about quality. It is that determinism is a feature. A tool that produces the same output from the same input is a tool you can build a pipeline on. Everything else is a very impressive slot machine.