Introduction

Tired of rolling the AI video dice and getting two characters side by side when you asked for them face to face? Topview just handed us a director's stage. Topview 3D Builder is now open source under Apache 2.0: you describe a shot to ChatGPT, Codex, Claude Code, Cursor or any agent with skills support (or hand it a reference image), and the agent builds an editable 3D set with characters, props and cameras. You lock the blocking and the camera move first, then send the renders to whatever video model you like.

Everything before the video step runs 100% locally, free, with no account and no upload. Projects live in a .topview-3d/ folder on your machine.

What Shipped

Topview announced the release on October 15 with the line "Stop rolling the AI lottery. Start building the shot." The repo at topviewai/topview-3d-builder is a TypeScript editor plus a Python CLI, and it was still getting commits on October 16 (plugin version 0.1.8).

Here's the toolkit in plain terms:

  • Agent-built scenes: the agent places built-in characters, poses and props, then adds story cameras and a bird's-eye overview camera.
  • Studio editor: a full 3D interface where every agent action has a matching hand-editing panel, so you can drag a character half a step without re-prompting everything.
  • Camera moves on a timeline: plan push-ins, pull-backs, rises, orbits and cuts before a single frame is generated.
  • Automatic checks: floating or clipping characters, spacing and framing get measured, so a bad layout is caught before you spend credits.
  • One scene, many cameras: render several angles from the same set, keeping positions consistent from shot to shot.

The Specs

  • License: code, schemas and docs under Apache 2.0; built-in assets under CC BY 4.0 (credit "Scene3D built-in assets © Topview, CC BY 4.0").
  • Built-in assets: 4 characters (child, youth, woman, man), 121 poses (sitting, standing, walking, fighting, crawling, reaching and more), primitive props, unlimited cameras.
  • Custom assets: import your own .glb/.gltf characters and props or pose .json files with asset import.
  • Requirements: Python 3.11+, Node.js 20.6+, and uv or pipx. The first render downloads a pinned Playwright Chromium once, then it's cached.
  • Render output: PNG frames, a contact sheet and a render.json record per run.
  • Recommended agent model: Topview suggests Qwen 3.8 Flash, which it says costs about 1/150 of GPT 6 Astra for a similar result.
  • Hosted option: on topview.ai, one 3D Builder scene build costs $0.10, with the video generation step running on Topview.

How to Try It Today

Claude Code users can add it as a plugin, then restart:

claude plugin marketplace add https://github.com/topviewai/topview-3d-builder.git
claude plugin install topview-3d-builder@topview-3d-builder

For Qoder or any agent with skills support, install the skill and pick topview-3d-cli:

npx skills add https://github.com/topviewai/topview-3d-builder.git

Cursor users clone it into ~/.cursor/plugins/local/topview-3d-builder, and ChatGPT and Codex users install it from the plugin page in the app. Then start a fresh session and give it a first shot:

Build a small café. Two people sit across a table from each other. Put the camera outside the window and slowly push in over four seconds. Render the first, middle and last frame.

Why It Matters for Creators

Text prompts are terrible at space. "Over-the-shoulder, then a top-down push-in" is a sentence video models love to ignore. A 3D blockout fixes composition, scale, eyelines and camera paths, while texture, lighting and performance still come from the video model. That split is exactly how real previs works, and it means fewer dead takes when you're paying per second of video.

It's also tidy for multi-shot work: one stage, several cameras, and the same spatial layout in every frame you feed into an image-to-video or reference-to-video model.

Original Source

Repo and docs: Topview 3D Builder on GitHub and the CLI reference.

Conclusion

Topview 3D Builder turns "please put the camera behind her" into an actual camera behind her. It's free, local, agent-native, and open enough to bend into your own pipeline. Block the scene, rehearse the move, then let the video model do the pretty part.

Got your previs frames? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=topview-3d-builder-open-source-ai-video-previs

—Aurelia ♡