CozyClay

AI video generators are good at producing a shot but bad at letting you plan one: you cannot easily decide where the camera sits, how the actors are blocked, or how a sequence of shots cuts together before you spend generation credits finding out. CozyClay borrows the previs step that film and animation productions have used for decades - staging a scene in rough 3D before committing to expensive production - and rebuilds it as a free, browser-based tool aimed squarely at handing the result to an AI video model afterward.

You block a scene with primitive 3D objects, control the camera with fly/pan/orbit navigation, and cut shots together on a timeline that supports dolly rails and standard camera moves. Character posing and motion can come from hand animation or from motion capture extracted out of ordinary video or photos via GVHMR, with sparse IK correction cleaning up the result. Once a shot is ready, CozyClay exports keyframe packs, MP4s, depth and normal passes, storyboards or OTIO cut lists, or feeds the scene straight to an AI video model such as Seedance, Kling, Veo or MiniMax H3. An MCP server and a `cclay live` CLI let an AI agent drive the same scene controls.

It fits small film, animation or content teams who want a previs stage before AI video generation instead of iterating blind in a text prompt, especially teams already comfortable with camera and blocking concepts from traditional production. Running via `npx cozyclay` with a prebuilt dist and no required build step lowers the barrier for anyone who just wants to try it, and the MCP/CLI support makes it a natural fit for teams that already orchestrate parts of their creative pipeline through agents rather than only through a GUI.

CozyClay is, in practice, a single-maintainer project - nearly every commit in its history comes from one author at NomaDamas, which is a real bus-factor risk for a tool a production might come to depend on. Some of its more advanced capabilities, motion capture and certain pose-generation paths, rely on GPU-based bridges (Kimodo, GVHMR, NVIDIA ARDY) that sit outside the core browser app and add real infrastructure requirements beyond just opening a webpage. The license is AGPL-3.0-or-later, so, as with OpenChatCut, a hosted service built on top would generally need to release its own source, worth checking before any commercial use beyond running it yourself.

The verdict: CozyClay is a genuinely useful bridge if your actual complaint about AI video tools is the lack of a planning stage - being able to block a scene and lock a camera move before spending generation credits is a real workflow improvement, not just a demo trick. The release cadence (weekly point releases through the date of this check) shows active, ongoing maintenance despite the small team. Treat the GPU-dependent motion-capture features as requiring more setup than the core 3D editor, and read the AGPL-3.0 terms before building anything commercial on top.