Clayducation

An AI video engine, built on Claude Code, that makes a narrated teaching video about a topic. New: soft 3D clay with a silent mascot who does the work. The classic 2D styles are still there.One agent checks the script before the voice is made.

Code draws every frame, without a video model, from 68 clay models of our own, and makes the music and the sound effects. The voice runs on your own computer, with or without a GPU.

Open the gallery page

Pipeline

What makes it different

A mascot who does the work	Formulas that act	Checked before render
The same silent mascot is in every video. He pulls, pushes and holds the things that the voice talks about.	When the voice says a term moves, it moves. When it says the mass cancels, the two terms are struck out and fly together.	The engine measures every frame. Claude reads strips of every movement like this one before the render can start.

Try it

In Claude Code, in the conversation where the topic came up:

Make a video explaining how the EKF fuses wheel odometry and the IMU.

Make a video on why computed torque control works.

Make a video on how backpropagation was discovered, and tell its history as a newsreel.

Two looks

New: clay, with a mascot

The default for every video: soft 3D clay, plain backgrounds and flat text, 94 models that we made in code, and one silent mascot who pulls, pushes and holds the things that the voice talks about. styles/clay.md has the mascot, his controls and every model.

Mahalanobis distance	PCA	k nearest neighbours
Bayes decision	LDA	Generalization

Frames from a series of six videos on image processing and computer vision.

He does the work with his own hands, and a different action can go in each part of a video. He carries, crouches to lift, sits, kneels, types, writes on a board, throws, knocks, presses, and grips a door handle, a lever or a wheel while it moves. He kicks a ball, jumps, hammers a nail in and pulls it out with the claw, sweeps dust with a broom, waters a plant that grows and flowers, and lies down to rest. He can also point, nod, shake his head and shrug. Each action is one helper in styles/library/clay.js. The helpers place his hands with the same contact measures as the rest of the video, so a hand that misses or goes into a thing is a fault before the render.

Classic: 29 flat styles

The first version's 2D styles, one style for each part of a video. Ask for them by name, for example "tell its history as a newsreel". They use the same workflow and the same checks as clay; see styles/classic/README.md.

Newsreel	Flat vector	Line boil
Pixel art	Risograph	Blueprint

Install

npx clayducation

The setup checks your machine, asks 4 questions (folder, voice, Manim, script agent), shows the plan, and installs only what you choose. It works without a GPU: then the Kokoro voice speaks and the CPU draws the frames.

Later, in Claude Code, /setup-clayducation changes the voice, the GPU or the Whisper model.

Clayducation had the name Kaleidoscope before November 2026. If you have the old install in ~/.claude/skills/kaleidoscope, run npx clayducation. It moves that folder to ~/.claude/skills/clayducation, keeps the voice and Manim environments, and updates it.

GPUs. An NVIDIA GPU speaks the voice, draws the frames and encodes the video. On a Mac with Apple silicon, the voice runs on Metal, Chrome draws with Metal and VideoToolbox encodes (the Mac path is new and not yet tested on a Mac; without a GPU the engine uses the CPU and makes the same frames, slower).

Whisper. make config --whisper large times the words and checks the voice with Whisper large-v3-turbo in place of the small base.en. On a test video it placed the cues within 0.01 s of where the speech starts, against 0.05 to 0.1 s with the small model. The installer chooses it on an NVIDIA GPU.

Disk space (estimates for a new install):

Install	About
Everything, NVIDIA GPU: Chatterbox voice, large Whisper, Manim	12 GB
Everything, Mac: Chatterbox on Metal, Manim with BasicTeX	7 GB (12 GB with the full MacTeX)
Smallest: Kokoro voice, no Manim	1.5 GB

This is the beginning of the README. Read the rest at the original. Read the full README ↗