← All Reviews

I Wrote a Video Using Only HTML and an AI Agent — Here's What Happened

heygen-com/hyperframes on GitHub
📦 heygen-com/hyperframes
49,718
Stars
🍴
4,568
Forks
🐛
138
Issues
🕐
7
Min Read
📝
900
Words
TypeScript Stable
View on GitHub →
ai animation ffmpeg framework gsap html mcp puppeteer rendering typescript

I'm going to be upfront: I was skeptical when I first saw HyperFrames. Write HTML. Render video. Built for agents. That tagline reads like every other buzzword sandwich in tech. But with 49,718 stars, 4,540 forks, and daily commits from a team that clearly knows what they're doing, I had to see what the fuss was about.

So I did.

What It Actually Does

HyperFrames sits in a gap I didn't know existed. If you want to generate video programmatically, your options are usually FFmpeg scripts (brutal), After Effects automation (expensive), or canvas-based web rendering (fragile). HyperFrames takes a different approach: you write real HTML and CSS, it renders that in a headless browser via Puppeteer, and then encodes the output to MP4 using FFmpeg.

The result is deterministic video from declarative web code. You describe what you want in HTML, define animations with CSS or GSAP, and the framework handles the rendering pipeline. No proprietary template language. No visual editor lock-in. Just HTML, CSS, and JavaScript.

Why It Matters Right Now

Three things make this more than a curiosity:

First, the AI agent integration is genuine, not bolted-on. HyperFrames ships 20 skills that AI coding agents (Claude Code, Cursor, Codex, Gemini CLI) can load on demand. The /hyperframes skill acts as a router — you describe what you want, and it picks the right workflow. This isn't a plugin that adds "AI features"; it's a framework that treats AI agents as first-class users.

Second, Heygen is a real company with real infrastructure behind it. This isn't some weekend project that'll go stale in six months. The commit history is relentless — Miguel Ángel and Vance have been pushing daily. v0.8.38 shipped the same day I was looking at the repo.

Third, the timing is right. Every developer and their team is experimenting with AI agents for content generation. HyperFrames gives those agents a concrete output format (video) instead of just text or images.

What Impressed Me

The monorepo architecture is genuinely well thought out. I count at least 12 packages: core, engine, producer, player, studio, shader-transitions, SDK, CLI, plus cloud deployments for GCP and AWS Lambda. The bundleToSingleHtml fix that shipped recently tells me they've actually wrestled with the real problem of making a standalone renderable bundle work reliably.

The skills system is smart too. Instead of forcing users to install everything, the router installs workflows on demand. npx hyperframes skills update gives you just the core set. It's lean by default and expands only when you need it.

The recent fixes — preserving nested sibling video layouts, inlining fonts and images in bundles, AAC correction headroom — these are the kinds of gritty, unglamorous problems that only surface when you actually render real video at scale. The team is clearly solving production problems, not just building demos.

Who Should Use This

If you're building automated video content at any scale — marketing clips, product demos, social media assets, educational content — and you'd rather write HTML than fiddle with keyframes in a video editor, this is worth your attention. The CLI workflow is solid, and the ability to pipe it into an AI agent workflow is a legitimate productivity multiplier.

If you're a solo developer or small team experimenting with agentic workflows, the skills system gives you a concrete use case that goes beyond "chat with your code."

Who Should Skip It

If you need real-time video editing or interactive video, this isn't it. HyperFrames produces deterministic, pre-rendered MP4s. You can't interact with the output.

If your team doesn't already work comfortably with TypeScript and Node.js, the learning curve is steeper than it looks. The Node.js >= 22 requirement alone rules out a lot of environments.

And if you're expecting this to be a free forever thing from a company that sells video AI products — you should read the fine print. Heygen has commercial products. HyperFrames is Apache 2.0 now, but the strategic direction could shift.

The Honest Concerns

132 open issues with this level of activity tells me the project is moving fast and some things are rough around the edges. The versioning (already at v0.8.38 after six months) suggests API instability. If you're building on this, pin your version and expect breaking changes.

The Puppeteer + FFmpeg dependency chain is heavy. Local renders will chew through memory and CPU. I'd be surprised if this runs comfortably on anything less than 8GB RAM.

And the "built for agents" angle, while clever, raises a question: how much of this framework is actually designed for humans writing HTML directly, versus being an interface layer for AI agents? The skills documentation is extensive, but the core developer experience for a human writing HTML and hitting npx hyperframes render isn't as clearly documented in what I've seen.

Verdict

HyperFrames is the most credible attempt I've seen at making HTML-to-video rendering actually production-ready, and the AI agent integration is the most thoughtful implementation of MCP/skills I've encountered. It's not polished enough for enterprise adoption without testing, and the pace of change means you'll need to stay current. But if you're in the video generation space or building agentic workflows that need video output, this is the repo to watch — and probably the one to try first.

The fact that it's Apache 2.0 with a real company behind it means you're not gambling your workflow on a ghost project.

Repository: https://github.com/heygen-com/hyperframes

// THE VERDICT
View heygen-com/hyperframes on GitHub →
Need help building with tools like this?
We build AI-powered applications and developer tools. 30+ years of engineering experience.
Get in Touch
hyperframeshtml-to-videoai-agentsffmpegtypescript
← Previous The dart-flutter-patterns Skill Has 246k Stars — But Is It Actually Worth Installing? Next → I Spent a Weekend With affaan-m's "Data Scraper Agent" Skill — Here's Whether It's Worth the Hype
← Back to All Reviews