EYamanS/chisel ? reverse-engineered prompt

Build me an MCP server for AI agents that can create real 3D models from simple shapes like boxes, spheres, cylinders, and cones, then combine them with union, subtract, intersect, and mirror. I want the agent to be able to build a model step by step, see a rendered preview after each change, and keep improving it until it looks right.

It should run headless, without needing a GPU or browser, and export finished models as glb or obj files. Please also make a simple web playground where I can try a few example models and see a live 3D view in the browser. Use the same modeling engine for both the MCP server and the web app, and make sure the scene can be reset, edited, and exported easily. If you need to check current docs for MCP or export formats, look them up online first.