vork, the Editorless 3D Engine: Scene Creation for AI Agents

Voyager’s vork, developed for AI agents, is presented as a 3D/VFX engine in which scenes are defined in code rather than through an editor interface. Reviewed in a user experience post published on 14 November 2026, the tool is in research preview and is available for free. Users create scenes with JSON, Python or TypeScript, then use the vork render command to produce images or video. The author of the post tested this approach—which relies on code and render output rather than an interface—in simulations at different scales.

Code and GPU instead of an editor

According to the post, vork’s core is written in Rust and uses Vulkan and Metal. The tool is designed to run on Windows, Linux and macOS. Materials, cameras and physics simulations can be added using small Slang programs that run on the GPU. The intended scene creation and refinement loop involves an AI agent writing code, inspecting the render output and revising the code.

This setup does not replace the visual editor experience found in traditional 3D applications. The author of the user experience post says that rendering and checking the result after every change makes fine-tuning harder; working on camera framing and backgrounds in particular can be cumbersome. As a result, vork appears to be designed less for artists who want to continually edit scenes on screen and more for tasks that can be described in code and are suited to automation.

Simulation and render times in user tests

The review reports that an example involving one million birds, rendered at 1080p over 192 frames on a single NVIDIA L4, took 97 seconds. This is the result of the official example cited in the post; the same time is not guaranteed on different hardware or with different scenes.

In the author’s own tests, simulating and rendering one million marbles took about four minutes on an RTX 5070 Ti for a 5.5-second video at 1080p. An animation featuring 93,000 zombie characters, converted into 24-frame shapes in Blender, took about 9 minutes and 20 seconds to complete 168 frames at 1080p. In a test adding flashlight illumination to live-action footage, rendering 120 frames at 1080p took about a minute; compositing the render with the footage took about two minutes. These are results reported by the author, not an independent product test.

What it means for architecture and visualization workflows

For architecture firms and archviz artists, vork’s potential uses include scenes with many repeated objects, physics simulations on the GPU or animated scenes that can be generated with code. However, the source does not specify direct compatibility with architectural project files or ready-made integrations with applications such as Revit or Blender. The tool should therefore be viewed not as a replacement for existing modeling and visualization software, but as a separate render engine to try for specific tasks.

GPU memory may be an important hardware limitation. The reviewer says vork can slow down noticeably when other applications are using VRAM, and that they closed Blender and ComfyUI before running it. This highlights the importance of managing shared hardware resources and scheduling render jobs for complex scenes. As for licensing, the current preview version is reportedly free and open to commercial use, although paid versions may be released in the future. Teams should check the latest license terms before using it.