Topview open-sources 3D Builder so agents can stage editable scenes before AI video

What Topview open-sourced on 15 October

On 15 October 2026, Topview published a DEV Community post under its TopviewAI account titled “From One Prompt to an Editable 3D Scene: We Open-Sourced a Local-First Agent Workflow.” The same day, the public repository topviewai/topview-3d-builder and the PyPI package topview-3d-cli 0.1.7 were the dated objects a reader can check: the wheel’s upload timestamp on PyPI is 2026-10-15T05:34:09Z, and the repository’s README and CLI contract document the install path, agent plugins and local project format.

The pitch is narrow. Describing a camera angle in text is easy; getting two characters, a table and an over-the-shoulder camera into the arrangement you meant is harder when every mistake costs another video generation. Topview’s answer is a staging step that lives on your computer: place characters and simple props, choose poses, position cameras, inspect framing, and render reference images that can inform a later AI-video workflow. Compatible coding agents drive that scene through a local CLI and an agent skill, starting from natural-language requests.

This is not a claim that Topview invented 3D previs. It is a claim that the company published a concrete open-source stack—skill manifests, a Python CLI and a local Studio—so an agent can build and revise a scene without uploading that project to Topview. The product page at topview.ai/3d-builder still describes a cloud plugin path that can charge for hosted staging; the open-source path documented in the DEV post and README is the account-free local route.

Why an editable 3D scene sits before the video model

Topview separates two decisions that generative video usually collapses into one prompt. Spatial intent covers where people and objects sit, which character is the subject, and where the camera points. Visual generation covers how the finished footage should look and feel. In the company’s example, an agent is asked to block a dialogue scene with two characters across a table, put a camera behind the first character aimed at the second, render a reference view, and keep the scene editable so blocking and camera can change afterward.

That framing sits next to other recent tools that try to give agents something more structured than a chat reply. We have covered Kling 4.0’s full availability for longer AI video clips and Skywork’s Matrix Game 4 open neural renderer; Topview’s bet is earlier in the pipeline. Fix the geometry first, then spend generation credits. The DEV post is careful on the next step: a reference frame does not guarantee that a downstream video model will preserve identity, composition or movement.

The useful part, Topview argues, is editing after the first render. If faces overlap, a table cuts the frame, or the camera is too high, you reposition characters, swap a static pose, move a camera, and render again from the same scene state—rather than regenerating an entire clip to learn whether the angle works. The scene is treated as software state, not as an unrecorded prompt retry.

What actually runs on your machine

The repository is more than a prompt wrapper. Per the README and CLI contract, three pieces sit together: agent skill and plugin manifests for documented agents; a Python CLI named topview-3d-cli that owns the project, validates structured edits, manages assets and orchestrates rendering; and a local Studio plus Node/Playwright renderer that uses Chromium to produce PNG reference frames.

Projects live in a .topview-3d/ directory inside a local folder. The CLI records a scene sequence, validates operations, and is designed so a failed write does not partially apply. Doctor checks require Python 3.11 or newer and Node.js 20.6 or newer, plus a one-time Playwright/Chromium setup via browser ensure. Core scene editing and rendering are local; installing packages and any cloud-hosted coding agent still need network access, which the DEV post states plainly.

Studio open starts a local server on 127.0.0.1:3002 when the port is free and reports a URL; it does not claim to open a browser for you. Render writes frame PNGs, an optional contact sheet and render.json with run id, resolution, camera, scene sequence, document SHA-256 and CLI versions. Chromium is configured to block requests outside the local server unless an optional public asset base is set—another local-first detail in the CLI contract.

How an agent builds and checks a scene

A typical CLI path starts with project init, project status, then studio open. Asset list and camera presets expose the limited built-in catalog. Agents apply changes through node batch and pose batch commands that accept JSON specs, optional dry-run flags, and an expectedSceneSequence guard against stale writes. Evaluate runs numerical checks—framing, origins in frustum, ground penetration, non-finite transforms—without opening a browser. Inspect views combines those checks with renders.

Batch size limits are explicit: node batch accepts 1 to 64 staged changes compiled into one atomic batch; pose batch accepts 1 to 16 characters, each once. A change that cannot apply fails with a named error such as NODE_BATCH_REJECTED or ASSET_NOT_FOUND rather than a silent partial scene. That matters when an agent proposes edits: the skill can check or dry-run before saving.

Install paths differ by agent. The README documents a ChatGPT/Codex plugin page, Claude Code marketplace commands, a Cursor local-plugin clone, and npx skills add for agents with skills support. Topview recommends Qwen 3.8 Flash for cost, saying the same task costs about 1/150 of GPT-6 Astra on its internal comparison; Opus 5.5 and Astra appear as stronger but more expensive options. Those model ratings are the vendor’s, not an independent bake-off—similar caution applies to other agent-tool launches we have covered, such as VoidZero’s Void host for vibe-coded apps.

Built-in assets, licenses and recommended models

Out of the box the repository ships four characters (child, youth, woman, man), 121 static poses across sitting, standing, walking, fighting, crawling, reaching and related categories, simple primitives for tables and walls, and as many static cameras as you add. NOTICE and the README split licensing cleanly: code, schemas and documentation under Apache-2.0; built-in characters, poses and specified assets under CC BY 4.0 with the attribution line “Scene3D built-in assets © Topview, CC BY 4.0.” Editor fonts ship under the SIL Open Font License 1.1.

There is no built-in walk-cycle or full-body motion library in this release. Position and rotation keyframes can suggest movement; they are not a fully animated character rig. Phase 1, the CLI contract says, ships no motion library: a missing motion clip still renders with a warning, and the character keeps its base pose. Topview would rather state that limit than market the project as a full 3D animation suite.

Using the local builder does not incur a Topview account fee. Whether it saves money on final video generation depends on your workflow and model; the DEV post does not claim a measured cost-reduction percentage. Sending a render into Topview Canvas is described as a separate, optional signed-in path.

Where this helps, and where it stops

Topview lists three promising uses: previsualizing composition before an expensive shot; comparing camera options on an identical spatial setup; and making iterations explainable—“move the second character half a step left” has a spatial meaning that “make it more cinematic” does not. Ad and e-commerce teams are called out alongside AI film creators who need character positions to stay stable across shots.

The hard stop is generative fidelity. A staging plan is not a finished clip. Model behavior, conditioning controls and identity preservation vary. Topview’s own language treats the PNG as evidence of intent, not as a contract with the video model. Teams that already fight inconsistent character placement across takes are the clearest fit; teams that only need a single stylized loop may find the extra tool chain heavier than another prompt.

What is still early

The project is early. Stars on the public repository were still in the dozens when we checked on 16 October 2026, and the DEV post asks for issues that name the agent, operating system, error and a minimal reproduction with private data removed. Release notes on 15 October also include a 0.1.8 pin for skills and plugins in the commit history; confirm the current CLI version before you publish install commands of your own.

Open questions for a second look: how often agents produce valid batches without human cleanup; how well reference frames transfer into Topview’s own video path versus third-party models; whether the asset catalog grows beyond four characters and static poses; and whether independent creators find the Studio fast enough to beat a storyboard scribble. Until those answers land, treat 15 October as the open-source availability date for a local staging tool—not as proof that previs solves generative video’s spatial lottery.