Some time ago, I became curious about the Roblox platform, observing how difficult it was for a beginner to generate structures. I began spending time integrating procedural generation and CAD using the Roblox Studio platform as a form of experimentation.
From there, the project kept expanding into a browser-based engine capable of generating truly structured buildings and worlds: walls, floors, roofs, windows, doors, rooms, and stairs, as well as terrain, caves, water, vegetation, lighting, and voxel environments.
The next phase involved making the engine work by representing rooms, openings, internal paths, stairs, and other constraints; this also allows the same basic structure to be transformed into different representations.
One particular element was the addition of a new layer that I called Terrain CAD; the goal was to give organic volume to the building. Today, the layer allows the same architectural structure to be transformed into voxel terrain, preserving openings, passages, and spatial logic while elements like stairs and roofs remain CAD parts that adapt to the scale of the structure.
Everything had to integrate with the platform so that the user could generate and send the generated structures/maps directly to the Roblox Studio without any manual process. During this time, I worked extensively with the Roblox plugin until we finally got it approved, the integration of all these elements was complex.
The platform also recently integrated an experimental MCP layer. The idea was not to give the agent access to the Procedural Engine, but to have it work through the CAD plan generated by the WorldLoop app. The agent can then read the plan and the structural report, understand where rooms, openings, stairs, and blocked areas are located, identify layout issues, and directly modify the structure definition to correct them. The result is then returned to the browser, where it can be checked before being sent to Roblox Studio.
Today, a fairly complete product is testable, but it still requires some work.
I'd love to get feedback given its "particular" nature.
Some time ago, I became curious about the Roblox platform, observing how difficult it was for a beginner to generate structures. I began spending time integrating procedural generation and CAD using the Roblox Studio platform as a form of experimentation.
From there, the project kept expanding into a browser-based engine capable of generating truly structured buildings and worlds: walls, floors, roofs, windows, doors, rooms, and stairs, as well as terrain, caves, water, vegetation, lighting, and voxel environments.
The next phase involved making the engine work by representing rooms, openings, internal paths, stairs, and other constraints; this also allows the same basic structure to be transformed into different representations.
One particular element was the addition of a new layer that I called Terrain CAD; the goal was to give organic volume to the building. Today, the layer allows the same architectural structure to be transformed into voxel terrain, preserving openings, passages, and spatial logic while elements like stairs and roofs remain CAD parts that adapt to the scale of the structure.
Everything had to integrate with the platform so that the user could generate and send the generated structures/maps directly to the Roblox Studio without any manual process. During this time, I worked extensively with the Roblox plugin until we finally got it approved, the integration of all these elements was complex.
The platform also recently integrated an experimental MCP layer. The idea was not to give the agent access to the Procedural Engine, but to have it work through the CAD plan generated by the WorldLoop app. The agent can then read the plan and the structural report, understand where rooms, openings, stairs, and blocked areas are located, identify layout issues, and directly modify the structure definition to correct them. The result is then returned to the browser, where it can be checked before being sent to Roblox Studio. Today, a fairly complete product is testable, but it still requires some work. I'd love to get feedback given its "particular" nature.