NomaDamas/CozyClay

Previs used to mean desktop 3D software and a long blockout session. CozyClay turns it into a browser tab: block a scene, generate motion, shoot camera moves, then export a keyframe pack that an AI video model can follow. If you already plan shots in Blender or similar tools and are happy doing that, treat this as an alternative, not a replacement.

The problem

The expensive step in AI video is the take. A prompt alone cannot describe where the cast stands, how they move, the lens direction and when the cut lands, so the model guesses and the result is re-rolled. The usual workaround is blocking the shot in a 3D tool, exporting reference frames by hand, and rewriting prompts until the footage matches. CozyClay collapses those steps into one workflow: the exported first/last frame, greybox clip, camera and prompt all come out of the same scene, so the model is asked to follow framing instead of inventing it.

What's different

The interesting design choice is the motion pipeline. Rather than requiring a rigged performance, the agent writes a motion spec and a deterministic compiler in the browser bakes every frame, with QA checking collisions, foot slide, reach and hand contact before you see the result. That means the movement you review is the movement exported — not a probabilistic generation you have to inspect frame by frame. The camera side is similarly concrete: shots sit on a timeline, each shot knows the limits of the targeted video model, and dolly rails are drawn in the scene.

The trade-off is that you are limited to what the spec/compiler pair supports; this is not freeform animation authoring. The AI-side surfaces are also broad — a chat panel with ChatGPT sign-in or your own provider key, an MCP server, and a terminal CLI over JSON. Live control from Claude Desktop, Claude Code, Cursor or any MCP client is convenient for fast iteration, but it also means the shot your agent produced needs an ordinary undo review; the good news is each agent edit is one undo step. Export formats lean toward a video pipeline (keyframe pack, mp4, depth and normal passes, contact sheet, OTIO), so the output is aimed at handing off to a generator, not at finishing the film inside the app.

Who should try it

A good fit is a solo creator or small team already paying per AI video take, with a clear idea of the blocking but no desire to learn a desktop 3D package. Storyboard mode, a row of panels with captions and an animatic, is a separate reason to open it even if no video generation follows. A poor fit is a character animator who needs precise hand-authored performance or anyone whose pipeline requires photoreal rendering inside the same tool; the greybox aesthetic and the compiler's motion vocabulary will feel like a cage rather than a shortcut.

Requirements and limitations

Running it is light: Node.js 22.19 or newer, a Chromium browser, and no GPU, Python or model download. That low bar is rare for 3D work and is the main reason the browser-tab framing holds. You can also skip local setup entirely by using the hosted version, though that raises its own question about where scenes and prompts are stored; the privacy doc covers anonymous telemetry and its opt-out.

Ongoing upkeep matters more than install. Video-model compatibility is stated per shot, so the capability limits of whatever model you target shape what the camera can do. Driving the studio through the agent panel, MCP server or CLI adds a provider key or a local MCP client, and agent-authored changes should be reviewed through undo rather than assumed correct. The AGPL-3.0-or-later license also deserves a read if you plan to build on the codebase commercially.

HotRepo's take

HotRepo has only two recorded daily-board days since 2026-10-24 (1,004 stars on 2026-10-24, +30; 1,005 on 2026-10-25, +28), for an 88-day-old repository. That is too little to judge whether interest persists. Before investing, check the camera-control guide for the specific AI video model you use, since the shot constraints are built around it.