~/codigoagente
AI agents

animate: the Claude Code skill that makes animated videos in code (free, no GPU)

animate is a Claude Code skill that makes animated videos in code: the agent writes the story, draws every frame on a <canvas>, composes the music with a JavaScript synthesizer and exports the MP4 with a headless browser and ffmpeg. There is no AI video model involved, so text stays crisp, every frame is reproducible and you need no GPU and pay nothing per generation. It is built for explainers, short stories, Shorts, Reels and TikToks in 9:16, 1:1, 4:5 or 16:9.

We use it for real: it is the engine behind codigoagente's vertical Shorts (notebook style, self-typing terminal, real screenshots with zooms). For this article we also cloned the repository (commit 7e5eb56, October 5, 2026), rendered its demos in a Claude Code cloud environment with no GPU and timed them.

The style gallery the skill itself builds with `tools/gallery.mjs` in under a second · [cth9191/animate](https://github.com/cth9191/animate) (MIT)

What animate is

  • Repository: github.com/cth9191/animate
  • License: MIT (copyright 2026 cth9191). Free to use, modify and use for client work.
  • Format: Claude Code plugin (/plugin install animate@animate) or a skill folder you copy to ~/.claude/skills/animate/.
  • Requirements: Node 18+, Playwright with Chromium and ffmpeg. For voice-overs, Python with faster-whisper.
  • Output: one self-contained index.html (open it in a browser to preview) and the final MP4 in renders/.

The core idea is that the video is a deterministic program: all randomness comes from a seeded generator (RNG(...)), never from Math.random() or the clock. Frame 240 is always the same frame, export can run in parallel, and a tool (tools/framehash.mjs) checks that a kit change did not move a single pixel of an approved piece.

What is in the repository

PartWhat it does
`styles/`7 styles, each with a `STYLE.md` (rules, palette, checklist), a drawing `kit.js` and a demo: cut-paper, crosshatch, riso, sketchbook, math, pixel and isometric
`grammar/`The story "grammar": 8 short-form formats, script rules (silence before the payoff, the end is the start changed) and a per-frame checklist
`kit/`Seeded randomness, a line that boils on 2s, cameras, shape-morph transitions, storyboard, synthesizer and a loudness stage for phones
`tools/`Build, stills, storyboard, export (WAV + stems + MP4), automated review, website capture, beat mapping for your track, voice and reference comparison
`craft.md`Everything learned building the pieces, as short rules
`examples/history-of-ai/`A complete 60 s cut-paper piece: from Turing (1950) to Claude Code (2025)

The most useful part for non-animators is the grammar. A history is told as a chronology joined by morphs (the old world closes in on one object, that object morphs into its counterpart in the new world, and the new world opens out of it). A "mission" cuts on the beat. A list is a catalogue. The agent picks the format from the plot, not at random.

How it works: three cheap check-ins before animating

The skill's first rule is no animating until you approve the storyboard. The flow has six steps:

  1. Intake. It asks only what it cannot infer: subject, formats, voice, style, hero, your music, your product. It shows you the style gallery.
  2. Story check. The one-line idea, the chosen format and a beat table with times, transitions and the role of sound. Every on-screen claim is web-checked and listed with its source.
  3. Look check. 2-4 full-size frames of the most different beats; you give a thumbs up per frame.
  4. Storyboard. Every beat as a key frame with its sound and the transition into the next. Approve or reject by panel number.
  5. Build and measured review. It animates, renders and runs tools/review.mjs: contact sheets per shot, cuts on the beat grid, text cut off or hidden by the phone UI, the loudest moment after the silence, and loudness in LUFS.
  6. Delivery. The MP4 with the check numbers, what was fixed and the weakest shot.

Your first video is one line away:

snippettext
/animate a 45-second explainer of how DNS works, math style, 9:16

Pieces are created in your project under pieces/<name>/.

Frames at 2 s and 6 s of the math style demo, rendered in the cloud with `tools/still.mjs` in 0.6 s · codigoagente

Step-by-step install

Dependencies first, on Windows, macOS or Linux:

terminalsh
npm i -g playwright && npx playwright install chromium
ffmpeg -version

If ffmpeg is missing, install it (winget install Gyan.FFmpeg on Windows, brew install ffmpeg on macOS, sudo apt install ffmpeg on Ubuntu). Then, inside Claude Code:

snippettext
/plugin marketplace add cth9191/animate
/plugin install animate@animate

If you would rather skip the plugin system, copy plugins/animate/skills/animate/ from the repository to ~/.claude/skills/animate/.

To check everything works without spending tokens, build and export one of the demos:

terminalsh
SKILL=~/.claude/skills/animate
cp -r $SKILL/styles/math/demo test-math
node $SKILL/tools/build.mjs test-math
node $SKILL/tools/still.mjs test-math 2,4 frames.png --scale 0.5
node $SKILL/tools/export.mjs test-math --share

In our test (a 4-core cloud container, no GPU) the build took 0.06 s, the stills 0.6 s and the full export of the 6 s demo (144 frames, audio, stems, final.mp4 and share.mp4) 9.2 s. Open index.html in a browser to preview: a click plays the score and ?t=12.5 freezes a frame.

Voice-over and music

The voice is the clock. With a voice, each beat starts 0.3 s after the previous sentence ends and the big moments land on specific words. The script lives in voice/script.json, one line per beat. You have three options for the takes:

  • The ElevenLabs connector in Claude (Settings → Connectors): the agent finds 2-3 narrators, generates one line with each, you pick, and it generates the rest. It costs about one credit per character and tells you the total first.
  • Your own recording, one file per line.
  • A free local voice saved as voice/<id>.mp3. For codigoagente's Shorts we use Supertonic 3 (MIT code, OpenRAIL-M model that allows commercial use), which runs on a CPU.

Then node tools/voice.mjs pieces/<name> listens to the audio with faster-whisper locally, gets the start and end of every word, and the animation follows. Change a line, rerun it, and the picture re-times itself.

One detail the README leaves out: voice.mjs defaults to Whisper's small.en model, which is English-only. For any other language pass --model small, or the word timings will be off.

With your own track, tools/beats.mjs detects tempo, bars and the big hits, cuts on the song's beats and lands the payoff on its loudest hit. Without music, the skill composes the score in code on a 120 BPM grid.

Real case: codigoagente's Shorts

We picked animate as the engine for the blog's vertical videos. Every Short follows hook → how the repo solves it → the full guide on the blog, in the sketchbook style with our own kit on top (self-typing terminal, real screenshots taped to the page, zooms where the information is) and the same voice every time.

Frames from the 27 s Short about the grill-me skill: typed terminal, real command and a real screenshot with a zoom · codigoagente

What the pilot taught us:

  • Handwritten fonts are not included. The skill uses system fonts with fallbacks; for the notebook look we installed Caveat and Permanent Marker (Google Fonts, OFL license) and aliased them in fontconfig.
  • Everything runs in the cloud. Node, Playwright, ffmpeg and the local voice work in Claude Code's remote container, so the video gets made with your PC switched off.
  • Synthetic voices mispronounce English names. For a Spanish voice to say them right, write the script the way it sounds ("Clod Cod" for Claude Code). The on-screen text stays correctly spelled.

What it costs to make videos with animate

PartPriceRequired?
animate0 USD (MIT)Yes
Claude Codefrom Claude Pro (20 USD/month) or pay-per-token APIYes
Node, Playwright, Chromium, ffmpeg0 USDYes
faster-whisper to time the voice0 USD, runs on your CPUOnly with voice
Local voice (Supertonic 3, Kokoro) or your recording0 USDOne voice option
ElevenLabsper plan, ~1 credit per characterNo
Music0 USD if the skill composes it; yours, with its rightsNo
GPUNot needed: 2D canvas in ChromiumNo

The cheapest setup, and the one we use: Claude Pro + local voice + skill-composed score = 0 USD extra per video. The only real cost is agent tokens; a 30 s video with three check-ins fits in a normal Pro session, although we did not measure exact usage.

What the README does not tell you

  • It is for short pieces. The grammar's formats target 15-60 s. For a 10-minute tutorial an editor or HyperFrames makes more sense.
  • It does not film or generate realistic people. Everything is procedural drawing. If you need a presenter or camera-style shots, this is not the tool.
  • Your real product, yes. tools/capture.mjs screenshots your site and animates it inside the style; it also takes your logo files from assets/.
  • One HTML file. The build rejects images, audio, web fonts or external URLs inside the code. The only exceptions are your screenshots and your music, both opt-in.
  • The WebGL style is "coming". The README announces a ray-marched 3D style; it is not included in the commit we reviewed.

Frequently asked questions

Is it an AI video generator like Sora or Veo? No. The AI is Claude writing code; a canvas draws the video frame by frame. No video model, no per-second generation cost.

Does it work on Windows? Yes: the skill has font fallbacks for Windows, macOS and Linux, and the scratch voice (--scratch) uses SAPI on Windows and say on macOS. You need Node, Playwright and ffmpeg on your PATH.

Can I use it for client videos? Yes. The license is MIT. Check the rights of any music and voice you add yourself.

Can I create my own style? Yes. Give it references (a video, stills or a web page); it measures palette, line weight, texture and cut rhythm, draws frames side by side with your references and, once you approve, saves the style in styles/<name>/ for reuse. References stay on your machine.

How is it different from video-shotcraft or Bang Motion? video-shotcraft uses Remotion and focuses on 16:9 product promos. Bang Motion makes HTML motion graphics with GSAP. animate is the most "illustrated" of the three: characters, paper, ink, pixels, synthesized music and native vertical.

Conclusion

animate is the cheapest way we have tested to make animated videos with Claude Code: MIT license, no GPU, no per-generation fees, and a check-in flow that saves you from redoing work. If you are starting out, install the plugin, export the demo of the style you like and ask for a 20-second video about something you know well. To see how it fits with voice, captions and other tools, read how to make AI videos with Claude Code or Codex on a budget; and to study how a video you like was made before remaking it, see rea-video.

// source code

Repository in this article
The code this article covers is public: browse it, star it or clone it from here.

Help us grow

codigoagente is free, with no paywalls. If this article helped you, any of these keeps the blog going.

// newsletter

One email a week with the repos worth your time

The week's new articles, tested against real code. No spam; unsubscribe in one click.

AI agents

guide · The complete ECC guide for Claude Code · #8

AgentShield, ECC's scanner, checks CLAUDE.md, settings.json, MCP, hooks and agents. We ran it on a deliberately vulnerable setup and on ECC's own install.
  • Security and code review
  • Claude Code
read article
AI agents

guide · The complete ECC guide for Claude Code · #7

We measured ECC's context footprint with /ecc:context-budget --audit (about 12,500 tokens on the minimal profile) and how to trim it: model, compaction and MCP.
  • Claude Code
  • Skills for your agents
read article
AI agents

guide · The complete ECC guide for Claude Code · #6

What ECC's 28 hooks do, how GateGuard stops the first edit, how to save and resume sessions and how continuous learning instincts work, with a real test.
  • Claude Code
  • Skills for your agents
read article