Up to 50% off yearly plansSee yearly prices

HyperFrames:
HeyGen's HTML-to-Video Framework, with Real Examples

Updated · 7 min read

This guide is about HyperFrames from HeyGen, the open-source HTML-to-video framework at github.com/heygen-com/hyperframes. It starts with four real projects and the skills that made them: two render with HyperFrames, one can run inside it or in plain HTML, and one ships its own renderer for a rule HyperFrames cannot check, because the fastest way to understand HyperFrames is to see what a page turns into. Setup and the case for Remotion come after.

What is HyperFrames?

HyperFrames is a free, open-source framework from HeyGen that renders an HTML page to an MP4. You write the video as HTML and CSS. Each visible element carries data-start, data-duration and data-track-index, and the motion is a paused animation timeline, usually GSAP, registered on the page. The renderer seeks that timeline to each frame in headless Chrome and encodes the frames with FFmpeg, so the same input gives the same video.

Its README describes it as a framework for turning HTML, CSS, media, and seekable animations into deterministic MP4 videos, and the repository's tagline is "Write HTML. Render video. Built for agents." It is Apache 2.0, needs Node.js 22 or later and FFmpeg, and the GitHub repository had more than 84,000 stars on 15 November 2026.

Because a frame is just a web page, anything a browser can draw can be in it: type, SVG charts, product screens rebuilt as cards, maps, captions and clips you supply. What it does not do is film. People and places come from footage or a video model, placed in the page as a video layer.

Case 1: How does a UI card promo use HyperFrames?

AdAnt's UI Card Promo skill has the agent rebuild a product's key screens as animated cards in HTML and SVG and render them with HyperFrames. The video on its page is a 25-second promo for an invented newsletter tool, Quorrel: a hook number, a question typed into a search bar, the logo, three proof cards, a summary graph and an end card.

How the skill builds it:

  • Every word and number lives in one data file the owner approves. Each card label carries data-bind, and a script fails any text typed straight into the page.
  • Each of the eight beats is one section of the page, marked as a clip, with its own data-start and data-duration. The example pins HyperFrames 0.9.20 and GSAP 3.15.0.
  • npx hyperframes check runs HyperFrames' lint, layout, motion and WCAG contrast passes, plus the skill's own audit for unbound, small, clipped and out-of-frame text.

In our run the check passed both cuts with 22 of 22 text elements meeting WCAG AA contrast. Both cuts rendered in 30 to 60 seconds on an Apple M4 Pro laptop: 13.9 MB at 1920x1080 and 10 MB at 1080x1350, H.264 at 30 fps.

One trap we hit: HyperFrames 0.9.20 renders a font family it does not find named in index.html in a system face, though snapshots look right. Name each family in an inline style block in index.html.

Case 2: How does Brag turn a codebase into a launch video?

Brag is an open-source skill by Latent Spaces that reads the current project's code, plans a short launch video, and hands a composition brief to HyperFrames to build and render it. Its default length is 15 to 25 seconds, in landscape, vertical or square, with music and sound effects; narration is opt-in and uses Kokoro through HyperFrames.

The split of work is clear in its own instructions: Brag owns the angle, storyboard, tone and audio choices, and HyperFrames owns the composition, the animation timing and the render. Its gate before rendering is npx hyperframes check passing with zero errors.

Two things to know before you run it:

  • On Claude Opus 5.5 it switches to a lighter /brag-slim workflow that builds the video with whatever tools are on the machine instead of handing it to HyperFrames. Run /brag --full, or ask for narration with --voice, to keep the HyperFrames one.
  • The full workflow runs npx hyperframes, which fetches the latest release unless one is installed. AdAnt's installer pins Brag to one commit and suggests npm i -D hyperframes@0.8.225 in your project first.

The clip on its page is a 12-second vertical video at 1080x1920 from the Brag project itself. We have not rendered a Brag film of our own.

Case 3: How does motionmaxxing build and review a launch film?

Motionmaxxing is an open-source skill by Tejas Makwana for launch films and promos, built on its own seek-safe GSAP runtime, and usable in HyperFrames or plain HTML. The approach is the same as HyperFrames': every frame is a pure function of time, so a renderer can jump to any moment and capture it.

What sets it apart is the review. After a draft render, scripts score the file against six gates: the render exists and moves, there is no run of more than 3 flat frames, the end card holds for no more than 1.4 s, and the page has no corner labels, progress bars or similar slide chrome. Two gates, readable proof and one hero per frame, are judged by eye. Its instructions say its eases come from curves measured frame by frame from 53 professional recreations.

Its provider check looks for HyperFrames along with Chrome, FFmpeg and an ElevenLabs key. When its rules and a HyperFrames house style disagree, its instructions say the skill wins. A clone is about 46 MB, so AdAnt's version is a small installer pinned to one commit, and it lists what the project's scripts read and send, since voice and image generation spend the user's credits.

Case 4: When does a narrated explainer skip HyperFrames?

AdAnt's Narrated Explainer skill uses the same idea, a page seeked frame by frame, but ships its own small renderer instead of HyperFrames. It needs one thing HyperFrames does not check: every picture must land on the word in the voiceover that names it.

The script is a JSON file where every figure has a source and a quote. WhisperX times every word of the voice take, and scenes are cued to words, not seconds. The renderer draws frame n at n / fps seconds with Playwright and pipes it to FFmpeg, and the checker fails any beat that moves more than a frame plus 0.04 s off its word.

The example on its page is a 19.2-second explainer on Earth's water at 1280x720, which rendered in about 20 seconds on an Apple laptop. The skill points to the HyperFrames skill for any HTML page to MP4 without a voice.

The lesson for HyperFrames users: the seek-per-frame model is simple enough to own. Use HyperFrames for the render and its checks, and add your own check where your video has a rule it cannot know.

What do the four cases have in common?

All four write the video as a web page whose every frame depends only on time, then capture it. UI Card Promo and Brag's full workflow render with HyperFrames, motionmaxxing runs inside it or in plain HTML, and the narrated explainer owns its renderer.

VideoSkillLength and sizeHyperFrames' role
Product promo of animated UI cardsUI Card Promo25 s, 1920x1080 and 1080x1350Renders and checks both cuts
Launch video from a codebaseBrag15 to 25 s, landscape, vertical or squareBuilds and renders in the full workflow
Launch film with scripted reviewMotionmaxxingSet per filmOptional; its own GSAP runtime
Narrated explainer synced to wordsNarrated Explainer30 to 90 s, wide, vertical or feedNot used; its own Playwright renderer

How do you set up HyperFrames with Claude Code?

Install HeyGen's skills or plugin, ask for a video, then pin the version once the project matters. The steps from HeyGen's README and quickstart:

  1. Install Node.js 22 or later and FFmpeg.
  2. In Claude Code, add the plugin: claude plugin marketplace add heygen-com/hyperframes, then claude plugin install hyperframes@hyperframes. Or install the standalone skills in your project folder with npx skills add heygen-com/hyperframes and pick Core Skills.
  3. Ask for a video. HeyGen's suggested first prompt starts "Using /hyperframes, create a 10-second product intro" and adds a title, a background video and music.
  4. Preview with npx hyperframes preview and render with npx hyperframes render.

Without an agent, npx hyperframes init my-video scaffolds a project by hand. For every step with its checks and the common errors, see the HyperFrames tutorial.

Two habits from our skills: pin HyperFrames in package.json (our promo uses 0.9.20), since npx hyperframes otherwise fetches the latest release, and turn off its usage telemetry with npx hyperframes telemetry disable if you don't want it. AdAnt also offers HeyGen's official skill at a pinned commit as the free HyperFrames: HTML to Video download.

When should you pick Remotion instead?

Pick Remotion when your team already writes React, or when you render one data-driven template many times on a mature cloud renderer. Remotion writes video as React components; HyperFrames writes it as HTML. HeyGen's own comparison says plainly that Remotion is older, with more templates, tutorials and production history.

The other deciding factor is the license. HyperFrames is Apache 2.0. Remotion is free for individuals and companies of up to 3 people, and larger companies need a paid Company License.

For the setup, the license and what a Remotion render installs, see Remotion with AI. For the two side by side, including the license table and which of our skills use which, see Remotion vs HyperFrames.

Frequently asked questions

Is HyperFrames free?

The framework is. It is Apache 2.0, and its README says there are no per-render fees or commercial-use thresholds. HeyGen's Studio app is free too; its paid Pro plan adds HeyGen's video, image, voice and music models, and some official skills, such as Reverse Any Video, need it.

Is HyperFrames the same as HeyGen's AI avatars?

No. HyperFrames renders HTML that you or an agent write; it does not generate people. Its media-use skill can fetch an avatar or generated clip from HeyGen and place it in the page as one more layer.

Do I need to know React or video editing?

No. A HyperFrames video is an HTML file with timing in data attributes and a paused animation timeline. The agent writes it, and you can open the file in a browser to see what it wrote.

Which AI agents work with HyperFrames?

HeyGen's README names Claude Code, Codex, Cursor, Gemini CLI and IBM Bob, and says other agents that support skills work too. Claude Code can also install it as a plugin.

Can I move a Remotion project to HyperFrames?

HeyGen publishes a remotion-to-hyperframes skill for it. It ports compositions to timelines and clips, flags what it cannot translate, such as React state, and compares the two renders frame by frame.

The studio dispatch

One useful idea on AI-made ads, every other week. No fluff.