Lemo-Opuscar is a free, open-source Claude Code plugin that turns Claude Opus into a filmmaker by having it write the rendering code for an entire animated film instead of generating frames with a video model. It is for people who want stylized animated shorts and are willing to drive a coding agent from a terminal to get them: motion designers curious about generative art, developers who enjoy Canvas and WebGL sketches, and anyone who would rather have a film style they can re-run and edit than a one-off prompt result.
What it does
The plugin, built by a developer who goes by Lemomo under the org handle lemomo-ai, installs as a Claude Code "skill" and gives the agent a library of film styles plus a method for making new ones.
- 59 ready-made styles across 11 categories, each one a style prompt paired with a demo film. The categories include documentary and edit styles, hand-drawn and painting looks, East Asian traditions (shadow puppetry, Ukiyo-e, paper-cut), printmaking, and 3D and materials work.
- 56 minutes of demo film content in total, by the README's own count.
- A guided five-step workflow for designing a custom style rather than picking one off the shelf.
The flagship demo, "Opuscar 98," is a 6 minute 31 second film that walks through all 98 Academy Award Best Picture winners from 1927 to 2025, each redrawn in a different matching visual style. The project counts 77 of them in 2D, 4 in 2.5D and 17 in 3D. Those figures are the authors' own description of their demo, not an independent measurement, and the demo is the main evidence on offer that the approach works at length.
How it works
Nothing here comes out of a diffusion or video model. The agent writes rendering code, Canvas and WebGL, that draws every frame; it also composes the music and generates text-to-speech narration, and FFmpeg stitches the result into a finished film. The README states that for "Opuscar 98" every frame, every note and every cut was written in code by Claude Opus 5.5, which is the project's own claim about its own demo.
The practical consequence is that the output is a program, not a video blob. You can read the code, change a number, re-render and get the same thing back with one difference, and keep it in version control. The materials do not name which text-to-speech engine or music tooling the generated code calls, so expect to find that out when you run it.
Getting started
The repository is at github.com/lemomo-ai/lemo-opuscar, and code and docs are MIT licensed. Install is two Claude Code commands:
claude plugin marketplace add lemomo-ai/lemo-opuscar
claude plugin install lemo-opuscar@lemolab
What costs money is the intelligence, not the plugin. Running it requires Claude Code and access to Claude Opus, so you need a paid Claude plan or API credit; there is no local-model path described in the materials. FFmpeg does the final assembly, so it has to be present on your machine. Beyond that the project doesn't list hardware requirements: no operating system, GPU, VRAM or memory figures appear in the materials. The demo films and the style gallery live on the project homepage at lemomo-ai.github.io/lemo-opuscar, which is the fastest way to decide whether any of the 59 looks is what you want before spending tokens.
When to use it / when not
The main catch is cost visibility: you are paying a frontier model to write, debug and iterate on rendering code, and the materials publish no token count or dollar figure for a film of any given length, so your budget is unknown until you have run one yourself. The second limit is by design, not a bug: every pixel is drawn by code, so photoreal footage is out of reach.
Use it instead of Sora, Veo or Runway when you want a look you can reproduce exactly, hand-edit and diff. Don't use it if you need live-action-style footage, have no Claude Opus access, or want a graphical interface, because the whole thing runs through Claude Code.
Alternatives
Mainstream AI video tools generate pixels straight from a video model and give you a file. Lemo-Opuscar sits at the opposite end: zero video-model frames, deterministic code instead, closer to a generative-art sketch or a game cutscene than to a text-to-video clip. Compared with simply asking Claude to animate something for you, what you are buying here is the 59 worked styles, their demo films, and a defined workflow for building a sixtieth.
Take this seriously if you make short animated pieces and care about owning the thing that produced them. The honest pitch is not that it beats a video model at realism, because it does not try, but that a film becomes a readable, re-runnable program with a style you can name and reuse. If you have never touched Claude Code, or you need a predictable per-film cost before you start, watch the demos first and treat the project as a convincing proof of a method rather than a turnkey studio.