AI for CAD Tools
Can Claude generate production-ready CAD designs? We tested it. Here's what it can generate, what it can't, and what works better for real engineering.
·
⏱
9 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
Claude can generate text descriptions, OpenSCAD scripts, SVG outlines, and Fusion commands via MCP. It cannot generate production-ready parametric CAD in STEP, IGES, or native formats. It has no manufacturing constraint awareness, no geometric verification, and no B-rep understanding. For engineers who need actual parts, the smarter approach is finding existing validated designs in your vault. Leo AI's patented geometry search and engineering knowledge base help teams find and improve proven designs rather than generating geometry from scratch.
"Can Claude design a part for me?" This question shows up in engineering forums every week. The answer is more nuanced than most people expect - and significantly more limited than the demos on social media suggest.
Claude is genuinely impressive at many things. Generating production-ready CAD geometry is not one of them. But the situation is not black and white, and there are specific scenarios where Claude's generative capabilities provide real value to engineers - just not in the way most people imagine when they hear "generative CAD design."
This review breaks down exactly what Claude can and cannot generate, the practical limitations that matter for production engineering, and what alternative approaches actually work for teams that need to get parts designed and manufactured.
What Claude Can Generate: The Text and Code Layer
Claude's generative strengths for CAD-related work live in the text and code domain. Understanding this boundary is important because it separates genuinely useful capabilities from wishful thinking.
Text descriptions and specifications. Ask Claude to describe the design requirements for a mounting bracket that needs to support 50kg under vibration loading, and it will produce a well-organized specification. Envelope constraints, material recommendations, fastener patterns, surface finish requirements - Claude can generate detailed text that helps an engineer scope the design before opening CAD. This is legitimately useful for early-stage work.
OpenSCAD scripts. Claude can generate OpenSCAD code from text descriptions. Describe a part - "cylindrical housing, 80mm OD, 60mm ID, 100mm tall, with a flanged base and four M6 through-holes on a 70mm bolt circle" - and Claude will write functional OpenSCAD code that produces the geometry. The output is parametric (you can change dimensions in the script), and it renders as proper solid geometry.
SVG outlines and 2D profiles. For 2D geometry - gasket profiles, sheet metal flat patterns, laser cutting outlines - Claude can generate SVG code that represents the shape. These are useful as starting points, though they typically need cleanup before going to a cutting service.
Fusion 360 commands via MCP. Through the Model Context Protocol connector, Claude can send create commands to Fusion 360 - sketches, extrusions, fillets, holes. The Fusion API executes these commands and builds geometry in the modeling environment. This works for straightforward part creation, though Claude cannot see or verify the result.
IN PRACTICE
Leo basically bridges the gap we had and allows us to design better products, faster products, for our clients. We come up with better, more creative, and more efficient solutions than we did before.
"Leo basically bridges the gap we had and allows us to design better products, faster products, for our clients. We come up with better, more creative, and more efficient solutions than we did before." - Harel Oberman, CEO, Oberman Industrial Designs
The Pros: Where Claude's Generative Capabilities Add Value
Let's give credit where it is due. For specific use cases, Claude's ability to generate CAD-adjacent content saves real time.
Rapid concept sketching in code. When you are exploring multiple bracket configurations or enclosure layouts early in a project, generating OpenSCAD scripts through Claude is dramatically faster than modeling each option manually. You can iterate on ten variations in the time it would take to model two in SolidWorks. These are not production models, but they are useful for making early geometry decisions.
Automation scripting. Claude generates functional macros and automation scripts for CAD platforms. If you need to batch-export files, rename configurations, update custom properties across an assembly, or automate repetitive modeling operations, Claude writes the code. This is not "generative CAD design" in the traditional sense, but it is generative work that directly supports CAD workflows.
Parametric templates. For simple, repeated part families - spacers, standoffs, adapter plates, simple brackets - Claude can generate parametric OpenSCAD templates where you plug in key dimensions and get the geometry. Teams that frequently design variants of similar parts find this useful for fast first passes.
Design documentation. Claude generates thorough design rationale documents, tolerance analysis write-ups, material selection justifications, and engineering change order descriptions. For the documentation layer that surrounds every design, this is a genuine productivity gain.
The Cons: Where It Falls Apart for Real Engineering
Here is where the review gets real. The limitations are not edge cases - they are fundamental constraints that affect every production engineering scenario.
No production-ready output formats. Claude cannot generate STEP, IGES, Parasolid, or native CAD files (SolidWorks, CATIA, Creo, NX, Inventor). OpenSCAD scripts can be exported to STL or STEP through the OpenSCAD application, but the geometry quality is limited by OpenSCAD's CSG modeling approach. The resulting files lack proper feature trees, editable dimensions, design tables, or any of the parametric intelligence that production CAD requires.
No manufacturing constraint awareness. Claude does not understand draft angles for injection molding, minimum wall thickness for casting, tool access for machining, bend radius limitations for sheet metal, or undercut restrictions. It can discuss these concepts in text, but it cannot evaluate whether a generated geometry actually satisfies them. This is the gap that separates a 3D shape from a manufacturable part.
No geometric verification. When Claude generates geometry through OpenSCAD scripts or Fusion MCP commands, it cannot verify the result. It does not know if features intersect incorrectly, if wall thicknesses are too thin, if draft angles are sufficient, or if the part is even watertight. Every generated geometry requires manual inspection by an engineer before it is useful.
No assembly context. Real parts do not exist in isolation. They mate with other components, fit within envelopes, clear fastener heads, allow assembly sequences. Claude generates geometry without any awareness of the assembly context, mating conditions, or system-level constraints that govern whether a part actually works in its intended application.
No B-rep understanding. Claude operates entirely in the text domain. It has no concept of boundary representation - the surfaces, edges, vertices, and topological relationships that define solid geometry in CAD systems. This means it cannot evaluate, compare, or modify existing 3D geometry. It can only create from scratch using text-based instructions.
What "Generative CAD" Actually Means in Production Engineering
The term "generative CAD design" carries different meanings depending on who you ask. In the marketing world, it suggests typing a sentence and getting a finished part. In production engineering, it means something much more specific and much more constrained.
Production generative design is really about informed exploration within engineering constraints. You define loads, boundary conditions, material allowables, manufacturing process constraints, and design space limits. A solver explores the design space and proposes geometry that satisfies all constraints while optimizing for weight, stiffness, or cost. The output still requires engineer review, cleanup, and validation before manufacturing.
Claude does none of this. It generates geometry from text descriptions without constraint awareness, optimization targets, or engineering validation. The gap between "Claude generated a shape that looks like a bracket" and "this bracket meets all structural, manufacturing, and assembly requirements" is the gap where real engineering happens.
That gap is not going to close through prompt engineering or better code generation. It requires fundamentally different capabilities - the ability to read and understand 3D geometry, evaluate engineering constraints against that geometry, and validate designs against manufacturing processes.
The Alternative Approach: Finding Instead of Generating
Here is a perspective that changes the whole conversation. Most engineers reaching for generative CAD tools are trying to solve a problem that does not actually require generating new geometry.
The majority of parts designed in any engineering organization share significant similarity with parts that already exist in the company vault. Brackets, housings, adapters, mounting plates, spacers - most of these have been designed before, validated in production, and have established manufacturing processes. The reason engineers keep designing from scratch is not that existing parts do not exist. It is that finding them is nearly impossible with traditional PDM search tools.
Leo AI takes this approach. Instead of generating new geometry (which Leo does not do either), it helps engineers find existing proven designs and improve them through engineering knowledge. Leo holds 3 US patents for reading CAD geometry natively - B-rep data, feature trees, assembly relationships. You can search text-to-text, text-to-CAD, or CAD-to-CAD.
Describe what you need in natural language, or upload a model and find geometrically similar parts across your entire vault. Leo offers integrations with leading PDM and PLM platforms including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM. The parts it finds are real - they have revision histories, associated drawings, manufacturing data, and proven supply chains.
Leo's Large Mechanical Model, trained on over 1 million pages of engineering standards and technical references, then helps you evaluate and improve those found designs. Material selection guidance with cited sources. DFM feedback based on actual manufacturing constraints. Tolerance analysis grounded in standards. The combination of finding validated geometry and augmenting it with engineering intelligence covers most of what engineers are actually trying to accomplish when they ask an AI to "design a part."
FAQ
Find Proven Designs Instantly
Search your vault by description or geometry
Instead of generating new geometry from scratch, Leo AI finds existing validated parts in your PDM vault using natural language or CAD-to-CAD search. Patented geometry reading across all major PLM platforms.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
Find Proven Designs Instantly
Search your vault by description or geometry
Instead of generating new geometry from scratch, Leo AI finds existing validated parts in your PDM vault using natural language or CAD-to-CAD search. Patented geometry reading across all major PLM platforms.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams