Yuanhao Feng

Back

CubeMuse

Describe a building in natural language to generate a Minecraft blueprint you can view, edit, and export.

Building generation and editing#

CubeMuse is an AI building tool for Minecraft players that runs in the browser. Describe a building, and the AI searches for references, builds the structure, and checks the result. The interface shows its actions and the changes to the building as they happen.

Blueprints can be rotated for inspection or viewed layer by layer with the slice control. To make changes, enter a follow-up request such as “make the roof red” to edit the relevant part. Manual mode provides an inventory and hotbar for placing and editing blocks, with Ctrl+Z to undo changes.

Each blueprint includes a material checklist and a step-by-step building tutorial. New blocks are highlighted in blue at each step, making it easier to follow along in Minecraft. Blueprints can also be exported as .litematic files for Java Edition or .mcstructure files for Bedrock Edition.

Workshop#

The Workshop contains blueprints published by users, with filters for category, size, and tags. Each blueprint includes a 3D preview and building tutorial. Blueprints can be copied and edited to suit a different build.

Technical implementation#

CubeMuse connects the language model to a geometry engine through a custom JSON build-instruction language (DSL). The model produces the build plan, and a deterministic engine calculates the block positions. Instructions must conform to a JSON Schema and pass structural and semantic validation. When validation fails, the specific errors are returned to the model for correction.

The system retains two versions of the generation engine. V2 serves as a fixed baseline for evaluating later changes. V3 uses an agent with limits on token usage, runtime, and tool calls. When a limit is reached, it returns the best snapshot saved during the build.

During construction, V3 uses a voxel software renderer to capture images and inspect the current structure. Once the build is complete, a vision model reviews it from four views to determine whether further changes are needed. Both versions are tested against the same fixed set of building prompts, with quality scores, token usage, and latency recorded for comparison.

Usage notes#

The service is hosted outside China, so access from mainland China may be unreliable. Generating buildings uses Credits, and new accounts receive a free allowance.