agentcad is an open-source CAD tool for coding agents, consisting of a CLI and an MCP server, with the goal of enabling AI agents to directly design and output real CAD results. It uses build123d as the default modeling engine and is compatible with CadQuery.
Main capabilities include: generating, rendering, validating, and exporting CAD models, with support for STEP, STL, and GLB export formats; providing previews, geometry validation, and version difference comparison; each run returns metrics and rendered images, based on which agents can discover and correct their own errors, progressively generating printable geometry.
How to use: you can paste the prompt directly into Claude Code, Cursor, or any coding agent, and the agent will automatically install agentcad and design its first part. You can also install it yourself with the command pip install 'agentcad[mcp]', which requires Python 3.10 to 3.12. For more installation options, see the official documentation. For MCP integration, the configuration can be added to .mcp.json, applicable to Claude Code, Cursor, or Windsurf, with no need to pre-install, as uvx will cache the environment on first run. Running agentcad init will write the Claude Code skill and AGENTS.md block into the project, letting agents understand the workflow in subsequent sessions; for existing projects, the skill can also be added separately. Scripts require no imports, as build123d basic primitives, show_object, and agentcad editing helper functions are all pre-injected.
Use cases: suitable for scenarios that require coding agents to participate in CAD design, such as completing 3D-printed model design, part modeling, rapid iteration from prompt to printable part, and human-agent collaborative CAD workflows through AI agents. Officially showcased cases include using Meta's Muse agent to design a 3D-printable iPhone Duo bedside dock and sending STL and USDZ files to a phone in about four minutes; and using a Claude session to build an end-to-end 3D reconstruction of the 1903 Wright Flyer, containing 234 parts, driven by prompts and faithful to the Christman measured drawings.
Limitations: the material does not provide explanations of specific operating systems, hardware requirements, model complexity limits, or commercial licensing restrictions. The source code is hosted on GitHub, uses the Apache-2.0 license, and can be run locally.
