How MACU Studio actually works
This is the deep dive - the reference behind the friendly guide. The guide shows you which buttons to press; these pages explain what happens when you do. The render pipeline stage by stage, every model that gets loaded and where its weights sit on disk, where the prompts live and how to rewrite them, how to point Studio at your own models, how to put a Claude Code agent in the driver’s seat, and how an episode gets from your GPU to YouTube to this site.
A pipeline built to be driven by AI
MACU Studio is an unapologetically AI-native tool. The pictures are made by a diffusion video model. The voices are cloned and synthesized by a neural TTS. The subtitles are aligned by a speech-recognition model. The shot lists, the title-card copy, the sound-effect spotting, and the 48-language dubs are all proposed by a local large language model running on your own GPU. Even the show you make with it is, at its core, a stack of model outputs stitched together by a deterministic pipeline you can read top to bottom.
We built it this way on purpose, and we built it this way too - Studio itself was written hand-in-hand with Claude. So the most important page here isn’t the pipeline or the model list; it’s Connect an agent. Wiring a Claude Code agent into Studio turns “write the script, click eight times, fix the four things that came out wrong” into a conversation: “tighten the cold open, regenerate Ron’s second line slower, re-render the lemonade-stand b-roll, and ship it.”Everything in these docs - every prompt file, every config knob, every endpoint - is something an agent can read and operate on your behalf. That’s the payoff: the more of this you hand to the agent, the less of it you have to do yourself.
The reference, page by page
New to Studio entirely? Read the friendly guide first, or click through the live demo - a real, fully-rendered episode where every button works and nothing you do is saved.