Guide 48 · Tutorial

Make Videos with Muse: HeyGen, Hyperframes & Skills Explained

How AI agents turn scripts into finished videos: HeyGen's open-source Hyperframes renders HTML into MP4, skills teach agents to drive it, and Muse's tool-calling fits the pattern.

By Abhit Tomar · Updated November 16, 2026 · 15 min read

Illustration of a film strip dissolving into HTML code brackets and animation keyframes
Illustration: museaicodes

Short answer

HeyGen's Hyperframes is a free, open-source tool that renders HTML pages into video files; “skills” are reusable instruction packages that let AI agents operate tools like it. Together they form a pattern — describe a video in words, get an MP4 back — that matches how Meta's Muse works with tools.

Watch: the video-creation workflow in 25 seconds

The one-paragraph version

There are three pieces. Hyperframes is HeyGen's free, open-source renderer that turns an HTML page into a video file. Skills are instruction packages (usually a SKILL.md file) that teach an AI agent how to use a tool like Hyperframes. And Muse — Meta's personal AI agent — is built around exactly this idea: an agent that picks up tools and does work with them. Put together, you get a workflow where you describe a video in plain words and an agent writes, checks, and renders it into a finished MP4.

This isn't theory: the comparison reels on our videos page were produced with this exact approach — HTML compositions rendered to MP4, voiced over, and published. Below is how the pieces fit and what they cost.

What Hyperframes actually is

Hyperframes is an HTML-to-video framework released by HeyGen under the Apache 2.0 license — free to use and self-host. You author a video as an ordinary HTML file: timing lives in data-* attributes on the elements, and motion is a paused GSAP timeline the framework can seek to any frame. The renderer opens the page in headless Chrome, seeks it frame by frame, and stitches the frames into an MP4 with FFmpeg.

The payoff is determinism: the same HTML always produces the same video, no matter how many times you render it. No timeline scrubbing, no render lottery. The basic loop is three commands: “npx hyperframes init my-video” to scaffold a project, edit the HTML, then “npm run render” to get your MP4. You need Node.js 22 or newer and FFmpeg on your machine; a “doctor” command checks all of that for you.

Authoring in HTML changes what a video is. The composition is a text file, so it lives in version control like code: you can diff two versions, revert a bad edit, and reuse a scene across videos by copying a block. Timing is explicit rather than felt — every element carries its own data attributes, and the GSAP timeline underneath is paused and seekable, so the renderer can jump to any frame without playing through the video first. That is what makes the output deterministic, and it is also what makes the workflow agent-friendly: an agent that can write HTML can write a video, and it can check its own work by rendering a single frame before committing to the full render.

  • Kinetic typography, counters, charts, lower thirds, and animated explainers are its sweet spot.
  • It also handles slideshows, product-launch videos, and ports of Remotion compositions.
  • Paid extras exist (cloud rendering, share links), but nothing in the core workflow requires them.

What HeyGen brings to the table

HeyGen itself is an AI avatar video platform: pick a presenter avatar, type a script, choose a voice, and get a talking-head video back — via their app or API. Hyperframes is its open-source sibling for everything around the presenter: titles, graphics, transitions, and full motion-graphic scenes.

Think of it as a division of labor. The avatar carries the human presence; Hyperframes carries the design. An agent-driven workflow can use both: generate the presenter clip with HeyGen's API, render the motion graphics with Hyperframes, and composite them into one final video. HeyGen's avatar minutes run on their credit system, so check their current pricing before planning a series — the Hyperframes half, by contrast, costs nothing but your computer's time.

The avatar side is where the money goes. HeyGen’s presenter clips are generated on their servers and billed by the minute through a credit system, so a series of talking-head videos has a real running cost — check the current pricing before you plan a season of them, because the numbers change and the plan tiers decide the resolution and quality you get. The practical split is to spend avatar minutes on the presenter and let Hyperframes carry everything else: titles, lower thirds, charts, and transitions cost nothing but render time. That division keeps the paid part small and the designed part repeatable.

What “skills” means — and why Muse fits

A skill, in this world, is a package of instructions that teaches an AI agent to use a specific tool well: what commands exist, what the input contract looks like, what good output looks like, and what to check before finishing. Hyperframes ships around twenty of them, installable into Claude Code, Cursor, Gemini CLI, Codex, or any agent that reads the skills format with a single “npx skills add heygen-com/hyperframes” command. After that, you can describe a video in plain English and the agent writes, lints, and renders the HTML for you.

This is the same pattern Meta built Muse around. Reporting on Muse's launch describes the model as trained for zero-shot tool calling using CLIs and skills — the agent discovers tools, reads their instructions, and operates them. So “skills for Muse” isn't a separate app store you download from; it's the capability pattern itself. Any tool packaged as a skill — video rendering included — is the kind of thing an agent like Muse is designed to pick up and use.

The workflow: words in, video out

  1. Describe the video in plain words: topic, length, aspect ratio (1080×1920 for reels), and the beats it should hit.
  2. The agent drafts the composition — an HTML file with timed sections — and a voiceover script to match.
  3. It lints and validates the composition (Hyperframes has built-in checks for malformed timing and contrast problems) before rendering anything.
  4. Render to MP4. If the design uses a HeyGen avatar, that clip is generated in parallel via the API.
  5. Review the output, revise the HTML, re-render. Because the source is text, version two is an edit, not a reshoot.

What it costs and what you need

  • Hyperframes self-hosted: free (Apache 2.0). Cloud rendering and team extras are paid and optional.
  • HeyGen avatar minutes: paid credits — check HeyGen's current pricing; it changes.
  • Your machine: Node.js 22+, FFmpeg, and enough disk for renders.
  • An agent that reads skills (or patience with HTML/CSS, if you'd rather write compositions by hand).

Honest limits

This is a creator workflow, not a button inside the Muse app — you're assembling the pipeline yourself (or with an agent's help). Hyperframes assumes you're comfortable with HTML and CSS unless an agent is writing it for you. Avatar video quality varies by plan and input, and local rendering is CPU-bound: a long 1080p piece takes real minutes. And skills are only as good as their authors — a sloppy skill produces sloppy video, so check a skill's source before trusting it with your brand.

Rendering is also CPU-bound, which sets the pace of the whole workflow. A short reel renders in minutes; a long 1080p explainer takes considerably longer, and every revision costs another full render — so the habit that saves the most time is checking a single frame or a short section before rendering the whole piece. And because the pipeline is assembled by you, there is no single support desk when something breaks: the renderer, the avatar API, and the agent’s skill are three separate pieces from three separate places, and a failure in any one of them is yours to diagnose.

Video skills, questions answered

Is Hyperframes really free? The core is: Apache 2.0, self-hosted, no account needed. HeyGen sells optional hosted extras like cloud rendering on top, but the render pipeline itself costs nothing.

Do I need to know how to code? Not necessarily. The skills exist precisely so you can describe what you want in plain English and let an agent write the HTML. But you'll get better results if you can read the composition well enough to spot when the agent misunderstood you.

Can Meta's Muse do this for me directly? Muse is built for tool-calling with CLIs and skills, which is the right shape — but whether a particular skill is available inside Muse's environment depends on Meta's integrations, which keep evolving. Treat this guide as the pattern; check what's actually connected in your Muse app.

HeyGen or Hyperframes — which do I need? Different jobs. HeyGen gives you a presenter who talks; Hyperframes gives you everything designed around them. Most polished AI videos use both.

Where do I start? Install the renderer (“npm i -g hyperframes”), scaffold the example project, and render it unchanged once. Seeing one successful MP4 demystifies the whole pipeline faster than any documentation.

Keep exploring

  • Video walkthroughs — our own reels, built with the HTML-to-video approach described here.
  • Codes directory — featured and community Muse referral codes, if you're setting up your account to follow along.

A first render, step by step

Start small on purpose. Install the renderer globally, scaffold the example project, and render it unchanged — the goal of the first run is not a good video, it is proof that Node, FFmpeg, and the renderer all work on your machine. The doctor command checks the toolchain for you; if it complains, fix that before touching any HTML. Once the example renders, open the composition and change one thing — a headline, a colour, a duration — then render again and compare the two MP4s. That loop is the whole workflow in miniature: edit text, render, review. Only then is it worth describing a real video to an agent: give it the topic, the length, the aspect ratio, and the beats it should hit, and ask it to build the composition the same way, one section at a time. When the agent’s first draft renders, watch it all the way through before asking for changes — agents are good at the composition and bad at knowing which of their own choices look wrong.

Can Muse make videos for you today?

Not as a button in the app — at least not in the form this guide describes. What Muse has is the capability pattern: it is built to discover tools, read their instructions, and operate them, and a video pipeline packaged as a skill is exactly the kind of tool that pattern is for. Whether a particular skill is available inside your Muse app depends on Meta’s integrations, which keep changing — so the honest answer is to check what is actually connected in your app rather than assuming the pipeline exists there. The value of this guide is the pattern itself: describe a video in words, let an agent write the composition, render it deterministically, and revise by editing text. That pattern works today with the tools above, and it is the shape video-making inside an agent will keep taking.