Prompt2CAD

The honest pitch for Prompt2CAD is narrower than "AI that does CAD." It's an agent that produces editable, dimensioned parametric solids in real-world millimeters and gives you a parts list with stable IDs — not a mesh you can only scale and rotate. That distinction matters the moment you need to change a shelf count or export STEP into Fusion 360, SolidWorks, or FreeCAD, and it's what separates the tool from AI mesh generators that produce visual assets rather than buildable geometry. What's genuinely useful: the conversational refinement loop. Saying "make the counter 20cm longer" or "add a second workstation" updates the same model rather than starting over, which is the workflow interior designer Federica Lencioni described publicly in July 2026. The parametric controls — width, height, shelf count, door style — cover the change axes that actually come up in furniture and fit-out work. The self-checking loop, where the agent screenshots its own geometry from multiple angles and corrects visible problems, addresses the most common failure mode of AI-generated models. The parts list is the reason a workshop planner would care. A generated cabinet comes back as P-001 side panel ×2 at 770 × 574 × 18 mm, MDF primer base, eggshell white finish, with hardware like concealed hinges and T-bar handles kept as separate rows — downloadable as CSV, carrying a "verify before fabrication" warning. For estimating and material buying, that's the difference between a pretty render and a costed job. The 2D-to-3D path is the newest surface and worth understanding precisely: it opens native DWG or DXF linework at measured scale, preserves supported layers, wall runs, openings, blocks, and arcs, and builds parametric solids for each object the plan describes rather than extruding every outline into a slab. It does not import text, dimension annotations, or paper space, and undimensioned plans yield proportions rather than precision — the agent estimates category-typical sizes and asks you for counter heights, clearances, and panel thicknesses. Where it doesn't fit: tolerances, mates, and engineering drawings for precision assemblies; sculpted or organic forms outside parametric modeling; and any workflow that needs production-signed geometry with no human review. The vendor does not publicly disclose which underlying AI models power the agent, so model details cannot be verified from the site. Treat the output as a strong starting point, check dimensions on anything load-bearing, and don't skip the review step before you cut material.