What they did
Filip Hanasz writes up a test run by his cofounder Szymon, a hardware engineer, who wanted to see whether a general-purpose Claude agent could design a real device without a dedicated AI CAD tool. The target was an open-source copy of the Pocket, a small magnetic voice recorder: an ESP32 board, two microphones, a LiPo cell and a machined aluminium case. The opening instruction was to install whatever tools it needed and do as much as it could alone. The agent set up CadQuery and OpenSCAD, pulled real STEP models of the boards, wrote the enclosure as one parametric Python script that fails on any part clash, ran its own manufacturability review, drew the machining drawings with matplotlib, and rendered the case by running a Blender script headless on a cloud machine. For parts it used a sourcing MCP server to price the BOM at JLCPCB, Mouser and Digi-Key. Szymon made the product calls on thickness, materials and colours.
How it went
After a week there was a full design with STEP files, toleranced drawings, a wiring diagram, a BOM and an assembly guide. The case went from 15 mm to 12.6 mm thick. The self review caught microphones sitting 1 mm low, an unsupported button and countersinks breaking through the lid. The first BOM pass mismatched a battery to a $49 industrial part and returned buzzers for a vibration motor; switching to exact part numbers fixed it and flagged a discontinued microphone. Core electronics came to about $33 a unit, and a test order for two units went out for about $65. Nothing has been built, the firmware is only a spec, and machined parts are unquoted.
Worth knowing
The authors build the parts-sourcing MCP server used here, so the sourcing step comes from the vendor side. The post does not name the agent harness, only Claude. Having the CAD script check interference itself was what kept the model from approving geometry by eye, and BOM matching needed exact manufacturer part numbers plus a human checking matches.