CozyClay is a browser-based 3D previsualization studio — block a scene, pose the cast, author camera moves and cuts, then carry the same shots to an AI video model. It is built with Three.js and React Three Fiber, licensed AGPL-3.0, and created and maintained by Doyun at the NomaDamas organization. Public since 28 July 2026, it has reached 319 stars and 43 forks, with v1.5.0 released on 23 August 2026 — the fifth release, also published to npm the same day.

Installation is a single command: npx cozyclay downloads the built studio and opens it at 127.0.0.1:5180, with nothing to compile and no dependency tree to install. It requires Node.js 22+ and a Chromium-based browser. The studio ships seeded with a pre-generated motion clip, so scrubbing the timeline, driving cameras, and drawing a dolly rail work immediately on a laptop — a sensible choice for a tool whose heaviest feature is optional.

The editing model borrows muscle memory from conventional 3D packages rather than inventing its own. Primitives and set pieces move, rotate, and scale through a W/E/R gizmo, with grid snapping treated as a preference that Ctrl inverts mid-drag. A bird's-eye plan view drives 2D root waypoints for character paths. Camera work covers right-drag fly with WASD walking and Q/E craning, middle-drag pan, Alt+drag orbit, and F to frame. Notably, every scene mutation goes through one history store — a drag, a scrub, or an inspector edit is exactly one undo entry, and Esc cancels an in-flight drag and restores the pre-drag transform. That is a detail plenty of larger tools get wrong.

Motion generation runs through ARDY, and this is where the hardware requirement lives: an SSH-accessible NVIDIA machine, with first-time setup pulling a ~16.4 GB text-encoder stack onto that box. Users pose characters, export poses, sequence multi-phase motion as Prompt Blocks on a resizable timeline, send them to ARDY, and play the result back with sparse IK correction where generated motion needs fixing. The project also ships a token-free provisioning path for ARDY's text encoder — one command installs the same stack from public repositories pinned by commit and SHA-256, skipping the Hugging Face account and gated-model approval. That workflow is built with Meta Llama 3, and the encoder's base weights carry the Meta Llama 3 Community License.

The distinguishing feature is the bundled MCP server. Connecting Claude or any MCP client lets a director ask for a shot in plain language and watch the viewport move — placing the cast, framing "a low wide profile", generating multi-phase motion. With the editor open, tool calls drive the visible viewport; with no editor running, scene and project tools operate headless to block scenes, derive film vocabulary such as "wide shot · right profile · knee level · 24mm", render AI video prompts, and write .cclayproject files. Four tools — capture_frame, set_prompt_blocks, generate_motion, and apply_batch — require the live editor.

The limits are worth naming. AGPL-3.0 is a deliberate choice that will rule the project out for some commercial pipelines. Motion generation is gated behind GPU hardware most users will not have idle, so a large share of the tool's promise stays off by default. And with 4 contributors and five releases in under a month, this is early software moving fast. For solo filmmakers and small teams who want to previsualize shots before spending on AI video generation, that trade may be easy; for studios wanting a stable pipeline component, it is early.