Mirage gives Claude tools to create and edit 3D scenes within a single project

A client asks to move a kettle in an advertisement. It is a small change, but you want to preserve the lighting, composition and camera movement that already work. Mirage's Tesseract offers a straightforward way to handle such requests: an agent changes the parameters of a saved 3D scene, then the engine renders the frames again. For a team that regularly revises product videos, the value lies in making controlled changes.

A Telegram post by @cgevent illustrates this approach with a nighttime kitchen scene in the rain, allegedly built entirely through Claude without a video generation model. The sources reviewed here do not confirm the production details of that particular clip. The tool itself and its capabilities, however, are documented: Mirage describes creating geometry and configuring materials, lighting and cameras through an agent, while retaining an editable scene and rendering it using path tracing. 3D tools overview.

What the agent does and what the engine does

Tesseract has two main components: skills containing instructions for the agent, and the tsrct command-line utility. The agent receives a task, consults the documentation and builds a .tsrct project. Subsequent user feedback becomes changes to that same project. The repository provides the command for adding the skills: npx skills add mirage-hq/Tesseract; it requires Node.js. Repository and instructions.

The division of responsibilities here should be understood literally. Claude interprets a request such as “make the lighting warmer” and chooses changes to the scene. The renderer is responsible for producing the image. Path tracing calculates how light travels, including reflections and indirect illumination. Those calculations alone do not guarantee a convincing image: composition, materials and model quality still matter.

One possible use case for a small team is a short video featuring its own device. The team has a model of the enclosure and needs a slow camera orbit and several color variations. First, it should agree on the shape and viewing angle, then add materials and lighting. Following feedback that “the camera is too close,” the team can work on a specific camera path while preserving the project's other settings.

This also gives discussions a more concrete focus. Instead of a vague “try again,” feedback can identify an object, a point in the movement or a light source. But the accuracy of the execution must be checked in the result: having editable parameters does not mean the agent will correctly understand the artistic brief on its first attempt.

Where the promises reach their limits

The original post speaks of “absolute control” and “infinite editability.” Those claims are too strong for a production workflow. Being able to change a scene does not eliminate agent errors, import limitations or the time needed to render it again.

Mirage explicitly cautions that Blender import is partial, and an environment unsuitable for GPU rendering may render a scene considerably more slowly. Tesseract also has no built-in model that turns any text prompt into a finished, complex 3D model. A real product may require a prepared source object. These limitations are listed in the official overview.

The pages reviewed do not establish how long it took to create the kitchen in the post, how many iterations were needed or what hardware was used to render it. Conclusions about budget savings or replacing a 3D specialist are therefore premature. Even a successful demo video does not tell you how predictably the tool will handle your product range and requirements for detail.

How to test its usefulness for your own task

For an initial trial, I would choose one object and a short camera movement. I would define three edits in advance: change the surface color, reposition a light source and adjust the camera path. Then I would check whether the approved parts of the scene remain intact and how long each iteration takes.

For a developer, the interesting aspect is giving an agent access to a specialized creative engine. For a product manager or founder, it is the opportunity to test the cost of a series of revisions using their own material. Mirage describes the engine itself as free; agent costs are paid separately. Product usage terms.

During such a test, it is worth saving the project, the source models and the final video. If the next edit can indeed be carried out predictably, there is a basis for bringing the tool into production. That repeatable result is more useful to assess than the visual impact of a single frame.