AgentCAD hands CAD to AI agents, and the geometry holds

A new open-source tool called AgentCAD lets coding agents design, render, and export real CAD output. The project ships as an MCP server and CLI, runs locally with no signup required, and produces valid STEP, STL, GLB, and OBJ files that can go straight to a 3D printer or a web viewer. The demo is a single Claude session that built a 234-part reconstruction of the 1903 Wright Flyer from a prompt.

The project is less interesting for what it is than for what it implies. AgentCAD is one of the first tools that closes a loop between natural language intent and physical geometry. The agent writes bad geometry on the first try, the tool returns metrics and a render, and the agent reads both and fixes its own mistakes. That feedback loop — run, render, inspect, fix — is the mechanism that matters.

AgentCAD was built by a developer using the handle jdilla1277 and is licensed under Apache-2.0. The source lives on GitHub. It uses Python 3.10 through 3.12 and the build123d library for its geometry kernel. The MCP integration means it plugs into Claude Code, Cursor, Windsurf, or any coding agent that supports the Model Context Protocol. A single JSON entry in .mcp.json is all it takes to wire it up.

The tool returns structured metrics after every run: dimensions, volume, face count, and a validity flag. The agent sees a JSON object like {"command": "run", "status": "success", "metrics": {"dimensions": {"x": 6480.8, "y": 12293.6, "z": 1999.9}, "volume": 305738441.28, "face_count": 684, "is_valid": true}}. That is the tight feedback loop. The agent does not guess whether the geometry is watertight. It reads the validity flag and the topology report and fixes the script.

The project also ships with an installable agent skill on skills.sh (Vercel) and ClawHub (OpenClaw). That matters because it lowers the friction of adoption. A user runs npx skills add jdilla1277/agentcad-skill or clawhub install jdilla1277/agentcad and the agent knows what to do. The skill is the documentation that the agent reads.

The case study is the most concrete signal of capability. A single Claude session built the 1903 Wright Flyer reconstruction from the Christman measured drawings. 234 parts. All built with prompts. The tool did not require a human to babysit the geometry. The agent wrote the script, ran it, inspected the output, fixed the errors, and iterated until the geometry was valid.

That is the kind of demo that tends to be a parlor trick until you ask what it costs in practice. The Wright Flyer reconstruction is impressive, but it is a known artifact with published drawings. The real test is whether an agent can design a custom enclosure for a sensor package or a bracket for a robot arm from a description like “60x40x25mm interior, 2mm walls, 3mm fillet on vertical edges.” AgentCAD’s own examples include that exact prompt, and the tool returns a valid 30-face, 72-edge, 39cm³ geometry.

The tool is not alone in this space. A paper posted to arXiv in March 2025, titled “From Idea to CAD: A Language Model-Driven Multi-Agent System for Collaborative Design,” by Felix Ocker, Stefan Menzel, Ahmed Sadik, and Thiago Rios, proposes a multi-agent architecture that mirrors an engineering team. The system uses a Vision Language Model with agents for requirements engineering, CAD engineering, and vision-based quality assurance. The paper demonstrates iterative validation loops with the user and provides ablations showing the benefits of each component.

AgentCAD is simpler. It is a single tool that gives an agent the ability to execute CAD scripts and read the results. It does not require a multi-agent architecture or a VLM. It just requires a coding agent that can write Python with build123d primitives and parse JSON metrics. That simplicity is a feature. It means any team using Claude Code or Cursor can try it today without standing up a multi-agent pipeline.

The business implications are straightforward. Hardware iteration is slow because the feedback loop between design and physical validation is long. A designer sketches, sends to CAD, waits for the model, sends to the printer, waits for the part, inspects, and iterates. AgentCAD compresses the first half of that loop. The agent generates a candidate design in seconds, renders it, validates the geometry, and presents the result. The human inspects and either accepts or revises the prompt. The iteration happens at the speed of a text edit.

That compression matters most for small teams and solo hardware founders. A single person with a 3D printer and a coding agent can now iterate on enclosure designs, jigs, brackets, and simple mechanical parts without learning CAD software. The skill ceiling for physical prototyping drops. The bottleneck shifts from CAD proficiency to prompt engineering and geometric intuition.

The tool has limitations that the README does not hide. It supports Python 3.10 through 3.12. It requires the agent to write valid Python with build123d. Complex assemblies with moving parts or tolerances are not demonstrated. The Wright Flyer is a static reconstruction, not a functional mechanism. The tool validates geometry, not engineering. A valid STEP file can still be a bad part.

The open-source nature of the project matters. Anyone can fork it, extend it, or audit it. The Apache-2.0 license means commercial use is unrestricted. The tool runs locally, so no geometry leaves the machine. That combination of properties is rare in the AI tools space, where most agentic CAD experiments are either closed-source research demos or SaaS products with data residency questions.

The question AgentCAD forces is not whether agents can generate CAD geometry. They clearly can. The question is whether the feedback loop is tight enough to make the agent useful for real hardware iteration. The tool returns metrics and renders, but it does not simulate physics. It does not check for manufacturability beyond geometry validity. It does not run finite element analysis. The agent can tell you the part is watertight. It cannot tell you the part will survive a drop test.

That gap is where the next iteration of this category will focus. The paper by Ocker et al. already points in that direction with its vision-based quality assurance agent. The natural extension of AgentCAD is a tool that not only validates geometry but also runs basic structural analysis, checks for draft angles, and flags features that are hard to machine or print.

For now, AgentCAD is a concrete demonstration that the gap between natural language and physical objects is closing faster than most hardware engineers expect. The tool is free, open-source, and works with the agents that teams already use. The 234-part Wright Flyer is a demo. The phone stand with a 5mm lip is the real product.