Agentic CAD put to the test: Claude Opus builds a Raspberry Pi 4 housing

A while ago I wrote about GPT-6 Astra generating parametric CAD. This time I ran my own test: Anthropic's new Opus model. The experiment could be summed up as “I got humbled”. Short version: I was impressed by Claude Opus 5.5, but not where I expected to be.

The setup: MCP straight into the live CAD session

Claude Opus 5.5 was connected directly to my running Fusion session via MCP (Model Context Protocol). One prompt, as simple as it was demanding:

Build a housing for a Raspberry Pi 4. Bottom shell with cutouts for every I/O port, air vent slots, snap-fit lid, ABS. And source the Pi 4 STEP file yourself.

What the model did entirely on its own

Data sourcing:It hunted down a Pi 4 STEP file, couldn't find an official one, checked the community model's license, and asked before downloading. A model that asks license questions before loading data, unthinkable a year ago.

Measuring instead of guessing: It imported the board and measured holes, PCB thickness and connector heights from the real geometry. Not “a Pi 4 is roughly 85 × 56 mm”, but real dimensions from real data. It also hit three dead ends in the Fusion API, and worked around each one on its own.

The result: Bottom shell, cutouts, vents and snap-fit lid in one pass. When I asked for screws, it found the right heat-set inserts and screws and implemented them. Close to one-shot quality.

The one moment I had to step in

Exactly one: the model started moving hundreds of bodies one by one instead of moving the component. A performance and architecture mistake I spotted after a few minutes and corrected. Annoying, but not a systematic failure, the kind of silliness you forgive a junior as long as they learn.

The critical part: those snap-fits

The snap-fits looked perfectly fine. And that's exactly where the problem starts: in 3D printing, I'd expect them to break off on first assembly. There isn't enough material behind them, the neck is too thin for layer build-up. And the housing isn't ready for injection molding either: no draft angles anywhere.

Why? Because I never told it. I never communicated which manufacturing process it was designing for.

Same geometry, two completely different parts

That is the point that matters: same geometry, two completely different parts. Print or mold, that decision drives the whole design:

• 3D printing: snap-fits need material behind the neck, build orientation decides layer lines and fracture behaviour, wall thicknesses get optimised for print time.
• Injection molding: draft angles on every wall, gate locations and parting lines planned, ribs designed against wall-thickness sinks, shrinkage factored in.

The model has no way to guess that. Give it the process and the constraints, and agentic CAD gets seriously potent. Skip that, and you get a housing that looks finished and breaks on first assembly.

The model does the work now. Knowing what to check is still my job.

It handles the geometry work: STEP import, measurements, features, workarounds. What it cannot do: take responsibility. Evaluating draft angles, weighing tooling cost against volume, translating a design decision by decision into something manufacturable. That remains engineering work, and that is exactly the value a design partner provides.

But: the gap is closing. Systems improve, and checklists plus structured prompts increase the odds of landing closer to a finished design with less human interference. I know how I'll prompt my next Raspberry Pi benchmark: process, material, volume, tolerances, IP rating, into the first prompt.

Note: the video for this test (the full run in time-lapse) will be added here shortly.

The CAD prompt checklist for your next project

If you are experimenting with agentic CAD yourself, put this checklist into the prompt, these are the points that separate “looks finished” from “is finished”:

• Manufacturing process: 3D printing (FDM/SLA/SLS) or injection molding? Either way, a different geometry.
• Material + volume: ABS, PA12, PP? 10 units or 10,000?
• Tolerances: which fits must hold, which are cosmetic?
• Fastening: screws, heat-set inserts, snap-fits, adhesive, including material compatibility.
• Environment: IP rating, temperature range, UV, vibration, contact media.
• Predecessor data:a STEP file of the electronics with real dimensions, not “roughly Pi-4-sized”.

With those six points, agentic CAD turns from “impressive experiment” into “usable engineering partner”. The rest is experience you can neither download nor prompt, but you can buy it.

Conclusion

Agentic CAD is no longer hype, it is a tool. On the day you provide process and constraints, you get in one session what used to be a few days of design work. And that is exactly why my job is changing: less clicking geometry, more bringing in and verifying manufacturing knowledge. The model will get better, but “who owns the CAD data?” and “is this manufacturable?” are still your questions to answer.

Anton Steenken

B.Eng. · Hardware R&D Engineer · Gründer von engineer your idea