HyperFrames — The Open-Source Remotion Killer I Was Looking For

Shared by @futuretrees

Updated May 21, 2026

HyperFrames — The Open-Source Remotion Killer I Was Looking For

TL;DR

I'd been hunting an open-source alternative to Remotion that's actually built for agents — not just a library you bolt agent prompts onto, but one designed from day one for AI to scaffold, animate, and render. HyperFrames is it. Apache 2.0, HTML+CSS+GSAP, MP4 export via Puppeteer+FFmpeg, ships with skill packages for Claude Code, Cursor, Codex. From HeyGen (yeah, the AI video company). Free. No per-seat fees. No per-render fees. No team-size cliffs. This is the one.

What it actually is

Repo: https://github.com/heygen-com/hyperframes Tagline: "Write HTML. Render video. Built for agents." License: Apache 2.0 (real open source, OSI-approved — no Remotion-style source-available trap) Stack: HTML + CSS + GSAP, no build step. Puppeteer captures frames from headless Chrome, FFmpeg encodes to MP4. Codebase: 97.5% TypeScript.

The "crazy shit with frames" angle: compositions are HTML elements with data-* attributes that declare timing and tracks. The engine walks frame-by-frame through headless Chrome, snapshots, and pipes to FFmpeg. No React reconciler. No build-step. You can open the HTML in your browser and see the composition play.

Why this beats Remotion for my use case

Aspect HyperFrames Remotion
Authoring HTML + CSS + GSAP React TSX
Build step None Required
License Apache 2.0 Source-available, paid past 3-person teams
Agent-native Yes — skills ship for Claude/Cursor/Codex No — you bolt prompts on
Determinism Headless Chrome Headless Chrome (same idea)
Export MP4 via FFmpeg MP4 via FFmpeg

The agent angle is the part I sleep on at my peril. Every other tool I just reviewed (Remotion, Revideo, Motion Canvas, Manim) is a library you can teach an agent — but HyperFrames was built for agents. The CLI defaults to non-interactive. The skills packages teach Claude Code, Cursor, Codex, OpenCode, Hermes, etc. how to scaffold, animate, and render. That's a different design center.

The HTML-first authoring matters too. Every LLM on earth writes great HTML/CSS. Most of them are middling at React with custom timeline primitives. HyperFrames lets the agent stay in its native fluency zone.

Confirmation from the ecosystem

The nexu-io/open-design project (Claude Design alternative I also starred) explicitly lists "HyperFrames" as an output format alongside HTML/PDF/PPTX/MP4. That's a signal HyperFrames is becoming a primitive other agent tools build on — the same way FFmpeg is a primitive. Good sign.

Likes

  • Apache 2.0. Real OSS. No team-size cliff. No per-render meter. Critical for anything Ditto-scale where I want users auto-generating share videos.
  • HTML/CSS/GSAP. Every dev (and every LLM) already knows the stack. Zero new mental model.
  • Agent-native by design. Skill packages, non-interactive CLI defaults. The thing was born to be driven by a coding agent.
  • Deterministic frame capture. Same Puppeteer + headless-Chrome approach Remotion uses — but without the React reconciler in the loop.
  • HeyGen behind it. Real AI video company, not a side project. Likely to keep getting maintained.

Dislikes / open questions

  • Young repo. Way smaller community than Remotion's 47k stars / 3M npm. Less Stack Overflow gravity if you hit something weird.
  • GSAP for animation. Great library, but GSAP's commercial license has its own gotchas if you use the bonus plugins. Stock GSAP is free.
  • Render performance unknown. Puppeteer+FFmpeg is the standard pipeline, but I haven't benched it against Revideo's Skia-based or Remotion's optimized Lambda path. For high-volume short-form, this matters.
  • HTML+CSS animation ceiling. For anything fancier than CSS transitions and GSAP tweens (think Motion Canvas-style generator-based choreography), the model breaks down. Remotion/Revideo's TS-first approach is more powerful for complex timelines.

How my recommendation matrix changes

Use case New pick Was
Automated per-user share videos at scale HyperFrames Revideo
Agent-driven video pipelines (Ditto generating videos in-thread) HyperFrames (no good answer before)
Hand-crafted React-coded compositions Remotion or Revideo Remotion
Hand-crafted explainer with voiceover Motion Canvas Motion Canvas
Math / equation visualizations Manim Manim
Audio-reactive generative loops Hydra Hydra
Quick music video, no code Banger.show Banger.show

The take

HyperFrames replaces both Remotion and Revideo as my default for anything Ditto wants to do programmatically with video. Apache 2.0 license removes the legal-ops drag. HTML authoring removes the agent's "translate React" tax. Skill packages mean Claude Code can scaffold a composition and render it without me writing the integration glue.

The play for Ditto: build a "share-as-video" skill that uses HyperFrames under the hood. User asks "make a video out of my last week of memories" → agent scaffolds an HTML composition with the user's data → HyperFrames renders to MP4 → ship to share.heyditto.ai. Zero license drag. Zero per-render fees. Agent-native end-to-end.

This is the move.

Found by searching my GitHub stars after my previous landscape report missed it. Lesson learned: stars are a memory layer the seach didn't index — worth grepping next time.