OpenSCAD MCP Server

The OpenSCAD MCP Server bridges the gap between natural‑language or image prompts and tangible 3D geometry, enabling developers to embed parametric modeling into AI‑driven workflows. By exposing a standard Model Context Protocol interface, the server lets an assistant such as Claude request a 3D model from a textual description or even a single image, then hands the resulting CAD file back in a ready‑to‑print format. This eliminates the manual steps of drafting, refining, and exporting models that traditionally bottleneck rapid prototyping.

At its core, the server orchestrates a multi‑stage pipeline. First, it generates one or more synthetic views of the target object using either Google Gemini or Venice.ai image generation APIs. These images are then routed through an approval workflow where a human reviewer can approve or reject each view before reconstruction. Once approved, the images are fed into CUDA‑based Multi‑View Stereo (MVS) to produce a dense point cloud and mesh. Finally, an OpenSCAD code generator translates the mesh into parametric SCAD files, preserving editable variables and enabling downstream customization. The entire process is wrapped in a modular Python architecture that separates AI integration, remote CUDA processing, and OpenSCAD rendering, making it straightforward to swap or extend components.

Key capabilities include remote processing of GPU‑intensive tasks, so heavy reconstruction can run on a dedicated server within the LAN while the MCP client remains lightweight. The server also supports exporting models in several parametric formats—CSG, AMF, 3MF, and SCAD—ensuring compatibility with a wide range of CAD tools and 3D printers. An optional network printer discovery feature allows the assistant to hand off finished models directly to a connected 3D printer, closing the loop from concept to physical object.

Developers can leverage this MCP server in numerous scenarios: rapid prototyping of product concepts, educational tools that turn descriptive learning materials into interactive models, or even automated design generation for generative‑adversarial networks that require a tangible output. By integrating the server into an AI workflow, teams can iterate on design ideas in minutes rather than hours, while still maintaining the parametric flexibility that designers rely on for fine‑tuning. The server’s modular design and adherence to MCP standards make it a versatile addition to any AI‑augmented engineering stack.