Most AI video tools start with a prompt and end with a generated clip you cannot fully control. Lemo-Opuscar takes the opposite route. It is an open source toolbox of film styles for Claude Code that turns a structured prompt into a short film written entirely in code, using Three.js, Canvas, and WebGL, with no video model involved (GitHub). Every frame, every note, and every cut is code, which means the result is editable, repeatable, and pixel sharp by design.
Key Takeaways
Lemo-Opuscar offers 66 ready to use film styles, each with a style prompt and a demo film made entirely in code by Claude Opus 5.5.
It renders picture with Canvas and WebGL, scores music from free sample libraries, and narrates with text to speech. No video generation, no stock footage.
You install it as a Claude Code plugin, name a style, describe your story, and get a finished short film, usually in 30 to 60 minutes of agent work.
The trade off is real: it needs Claude Opus 5.5 for best results, plus Node 20+, ffmpeg 7+, and Python 3.11+, and it is slower than a one click video generator.
What Is Lemo-Opuscar?
Lemo-Opuscar is a project by Lemomo, built for Claude Code. Its description on GitHub reads: "A toolbox of film styles made in code for Claude Code: 66 ready to use, and your own. Each style is a style prompt plus a demo film made entirely in code by Claude Opus 5.5, with no video model." (GitHub). The project is public under the MIT licence, and at the time of writing it had passed 10,220 stars and 1,167 forks on GitHub.
The core idea is simple. Instead of asking a video model to hallucinate a scene, you ask a coding agent to direct a small studio. The agent reads three guides: one on directing (story, sound, rhythm, camera, performance), one on technique (frame by frame rendering, voice, music, mixing), and one style file that defines the look and sound. Then it writes the code, renders the frames, and assembles the film.
Watch the 30-second repo signal that first surfaced this project: Watch the 30-second repo signal.
How a Film Made of Code Actually Works
A Lemo-Opuscar film is a web page brought to life frame by frame. Picture comes from Canvas and WebGL pages, rendered one frame at a time. Music comes from original scores built on free sample libraries. Voice comes from text to speech narration. The finished video is assembled with ffmpeg, at a default output of 1920x1080 and 24 frames per second (GitHub).
One demo shows how concrete this gets. The Paper-cut Lightbox demo film, *The Last Night Train*, runs 51.0 seconds at 1920x1080 and 30 frames per second, using three.js 0.170 with WebGL, Canvas2D drawn paper layers, and headless Chrome frame capture (GitHub). Nothing in that pipeline is a generated video clip. If a shot looks wrong, you fix the code, not re-roll a prompt and hope.
That control is the point. Text stays sharp because it is drawn, not guessed. Data animations stay accurate because the numbers come from your script. A chart can move exactly on the beat, because the beat is in the code too.
66 Styles, From Ukiyo-e to Evidence Films
The library has grown fast, so older write ups quote older counts. A guide published on 30 September 2026 counted 43 styles (Open-source Projects), and a later review counted 59. The current README, verified on 1 November 2026, lists 66, after 7 new styles arrived on 29 October 2026, a jump independently noted the same week (Triffin').
The styles span hand drawn and painting looks, East Asian traditions, print and printmaking, graphic and type design, information and keynote visuals, cartoon and anime, games, cinema eras, and materials and 3D. Recent additions include Edit and Documentary styles cut from real footage, documents, and data, and Evidence Films, a set of 8 styles that film real material with every number sourced.
The showcase piece is OPUSCAR 98, a 6 minute 31 second film that walks one character through all 98 Best Picture winners from 1927 to 2025, each redrawn in a fitting style, with no video model (LemoLab). Its 98 styles break down as 77 in 2D, 4 in 2.5D, and 17 in 3D, according to the README. It is the clearest proof that this approach scales beyond a single short.
How to Make Your First Film
Installation is a Claude Code plugin, and it is the recommended route:
Run `claude plugin marketplace add lemomo-ai/lemo-opuscar`, then `claude plugin install lemo-opuscar@lemolab`.
On first use, the guides, tools, and style prompts download to `~/lemo-opuscar`.
Ask for a film in plain language, for example: "Make a 45-second film in the watercolor style about the coffee farm my family runs. Warm female narrator."
The agent asks a few questions up front: anything it cannot decide about your topic, whether you have your own voice, music, or other material, and whether you want to see a storyboard first. Say yes to the storyboard and it stops once to show key shots in the real style. Otherwise it goes straight to the finished film. You can also clone the repo instead of installing the plugin, in which case films go into `films/<name>/` inside the repo.
You can also make a style your own. Change a library style into a branded version, or start from a blank page and describe the look in your own words. The agent walks through the change one question at a time, shows one style frame first (side by side with the original when you adapted a library style), renders a 15 to 30 second demo film, and only saves the style when you are happy. Your own styles live in `~/lemo-my-styles/` and survive updates and uninstalls.
Code Films vs AI Video Generators
Aspect | Lemo-Opuscar | Typical AI video generator |
|---|---|---|
How the picture is made | Code: Canvas, Three.js, WebGL | A video model generates pixels |
Text and numbers | Sharp and exact, drawn from script | Often warped or hallucinated |
Editing a shot | Change the code and re-render | Re-roll the prompt, new result each time |
Speed | About 30 to 60 minutes per film | Seconds to minutes per clip |
Best for | Explainers, data stories, branded shorts | Photorealistic or cinematic clips |
That table also explains who should skip it. If you want a photorealistic scene of a person walking through rain, a video model is still the faster tool. If you want a data explainer, a product story, or a stylised short where every word on screen must be correct, code wins.
Limits Worth Knowing Before You Start
Lemo-Opuscar is honest about its constraints. The styles are tuned for Claude Opus 5.5, and other models may not reproduce them well. A film takes an agent about 30 to 60 minutes and a fair amount of tokens, so this is not a bulk content machine. You need Node 20+, ffmpeg 7+, and Python 3.11+ (or uv) installed; note that Ubuntu 24.04 ships ffmpeg 6.1, which is too old, so a static build is needed there.
It is also a young project, created on 26 September 2026, so styles, guides, and tooling are still moving. The fast growth from 43 to 66 styles in about a month is exciting, but it also means a style you use today may be refined tomorrow. For a related take on AI generated video, see how another open source tool turns existing video into a stylised remake: Soap2Soap: Turn Any Video Into a Pixar-Style Remake With Open Source AI.
Frequently Asked Questions
Do I need to know how to code to use Lemo-Opuscar?
No. You describe the film and pick a style, and Claude Code writes the code. Knowing basic code helps if you want to tweak a shot yourself, but the normal flow is prompt in, film out.
Is Lemo-Opuscar free?
Yes. It is open source under the MIT licence. You still pay for the Claude Code usage and tokens your films consume, and you need the tools listed above installed.
Can I use it with AI models other than Claude Opus 5.5?
You can try, but the README warns that styles are tuned for Opus 5.5 and other models may not reproduce them well. For reliable results, use the recommended model.
How long should my first film be?
Start short. The project suggests 30 to 45 seconds for a first film, which keeps the agent time, token cost, and review loop manageable while you learn how a style behaves.
The Bottom Line
Lemo-Opuscar does not try to beat video models at their own game. It changes the game: a film as a program, where style is a reusable prompt, text is drawn rather than guessed, and every shot can be edited like code. For explainers, data stories, and stylised shorts, that control is worth the slower pace. The library is already at 66 styles and growing, the showcase film proves the ceiling is high, and the plugin install makes the first film a one sentence request away.
Read this article in Indonesian: Read this article in Indonesian.