What Is AgentCAD and Why It Deserves Attention

AgentCAD is an MCP server and CLI tool that connects AI coding agents — like Claude Code, Cursor, Windsurf, or OpenAI Codex — with a professional CAD environment. It was created by developer James Dillard and drew immediate attention on Product Hunt: within the first week it gathered over 160 followers and ranked among the top-rated products of the day.

While common AI 3D tools (like Meshy or Tripo AI) generate models directly from text descriptions using neural networks, AgentCAD takes a different path. It leverages the programming abilities of AI agents — they write Python scripts using build123d and CadQuery libraries, while AgentCAD provides real-time feedback: it checks geometry, renders previews, and exports finished files. In other words, it's not "AI that magically creates 3D," but a clever bridge between coding agents and engineering tools.

How It Works in Practice

The process is surprisingly straightforward. You give the agent a prompt — for example, "design an electronics enclosure with a snap-fit lid, internal dimensions 60×40×25 mm, wall thickness 2 mm". The agent writes a build123d script, runs it through AgentCAD — and immediately sees the result: a four-view preview, geometric metrics (volume, face count, dimensions), and most importantly validation of whether the model is "watertight" and ready for printing.

If the geometry doesn't check out — for example, a wall is missing or the model isn't closed — the agent fixes the error itself and iterates the design. This eliminates the biggest pain point of AI-generated CAD: until now, agents wrote code blindly with no way to check their work. AgentCAD closes this loop: run → render → validate → fix. The result is a model that has passed automated quality control before a human even sees it.

Specific Features AgentCAD Offers

The tool covers the complete workflow from code to finished 3D file:

• Script execution — the agent writes code, AgentCAD runs it and generates a versioned STEP file with metrics
• Rendering — PNG previews from any angle, including GIF animations for all-around inspection
• Geometry validation — watertightness check, detection of loose edges and missing faces in under 100 ms
• Format export — STL for 3D printing, GLB for web viewers, OBJ for further editing
• Topology inspection — detailed report on model structure for debugging
• Diff tool — comparison of two design versions (metrics, outputs, parameters)
• Interactive viewer — display STEP/GLB files directly in the browser
• MCP integration — native support for Claude Code, Cursor, and Windsurf via .mcp.json

An interesting detail is that scripts require no imports — AgentCAD automatically inserts build123d primitives and helper functions, so the agent can write box = Box(10, 20, 5) right away. This dramatically reduces errors and simplifies the entire process.

Pricing and Availability

AgentCAD is completely free and open-source under the Apache 2.0 license. It runs locally, requires no registration and no cloud services. Installation is via pip: pip install agentcad[mcp]. It currently supports Python 3.10 through 3.12 (due to dependency on the OpenCascade library, which doesn't yet work on Python 3.13+).

For European users, this is good news — AgentCAD requires no subscription, no API keys, and no corporate licenses. All you need is a computer with Python and any coding agent, including free options like GitHub Copilot Free or local models.

Comparison with Alternatives

There are several tools at the intersection of AI and 3D modeling on the market. They differ in approach:

Meshy and Tripo AI generate 3D models directly from text descriptions or images using diffusion models. The results are visually impressive but often lack precise dimensions and aren't suitable for technical applications. Pricing ranges from a free tier to tens of dollars per month.

Autodesk Fusion with Claude connector — Anthropic recently announced linking Claude with Fusion 360. This is a proprietary solution requiring a Fusion license (starting at $70/month) and focuses on assisted modeling rather than fully autonomous design.

Traditional CAD software (FreeCAD, OpenSCAD) — requires manual modeling or script writing. AgentCAD adds a new layer here: an AI agent that writes scripts for you while also checking their output.

AgentCAD differs from all of them in that it is not a 3D model generator, but a tool for coding agents. It relies on the agent's ability to write correct code — and gives it feedback that would otherwise have to be provided by a human.

Practical Applications: From Prototypes to Production

On the AgentCAD website you'll find a gallery of models created by agents from a single prompt — an elegant vase, a rook chess piece, a phone stand, or an electronics box with a snap-fit lid. Each was created purely from a text description and passed geometry validation.

The most ambitious showcase is the reconstruction of the 1903 Wright Flyer — a complete model comprising 234 parts, faithful to the period measured drawings of Christman's plans. The entire model was created in a single conversation with Claude Code. While this is more of a showcase, it demonstrates potential for historical documentation, education, and rapid prototyping.

For the average user, AgentCAD makes sense primarily in three scenarios:

1. Custom 3D printing — design an enclosure, holder, or adapter in minutes instead of hours in CAD
2. Rapid prototyping — engineers and designers get a first iteration in a fraction of the time
3. Education — engineering students see how a parametric model is born from a description

Technical Background: build123d, CadQuery, and MCP

Under the hood of AgentCAD, two established open-source libraries do the work: build123d (the default runtime) and CadQuery (optional). Both are built on the OpenCascade geometry kernel — the same engine that powers professional tools like FreeCAD. This means AgentCAD outputs are compatible with industry standards.

A key architectural decision is the use of the MCP (Model Context Protocol). Thanks to it, AgentCAD behaves as a native editor extension — the agent "sees" it as a set of tools, not as an external program. This enables tight integration without needing to configure APIs or authentication.

Limitations and Things to Watch Out For

AgentCAD is a young project — the first public version was released in early June 2026 and the GitHub repository barely counts a dozen commits. Keep this in mind:

• Python 3.13 is not yet supported — due to dependency on OpenCascade bindings
• Output quality depends on the agent's capabilities — Claude Code performs excellently; less capable models may produce unusable results
• More complex geometry (organic shapes, freeform surfaces) is not a strong suit — the tool targets technical, parametric modeling
• No Czech localization exists — both documentation and interface are in English, which isn't a major obstacle for a technical tool

Nevertheless, the direction AgentCAD points to is significant. It falls into the trend of "vibe engineering" — the equivalent of "vibe coding" brought into the world of mechanical design. Coding agents are learning to interact with CAD tools and iteratively improve their outputs.

How to Get Started in Five Minutes

Installation is simple. Create a Python 3.12 virtual environment and run:

Then paste this prompt into Claude Code, Cursor, or another agent:

The agent will install the necessary dependencies itself, write the script, run AgentCAD, validate the geometry, and return the finished model along with previews. No clicking in CAD, no manual drawing.