CHICAGO — Darren Henry is a self-described CAD junkie with over three decades in the industry. He has served with SolidWorks and Onshape and now with PTC. We are talking serious CAD cred.
Darren took the stage at PTC NEXT Tuesday afternoon, respectfully after PTC’s new CEO, Neil Barua, with a specific CAD-related mission: separate signal from noise in the AI-and-CAD conversation.
Henry is well aware of all the AI-generated CAD models, MCP-to-CAD demos and breathless claims about disrupting the whole manufacturing. He’s not impressed.
“The models are simplistic,” he said. “They’re extremely fragile. They don’t have design intent.”
Take a washer, for example. Sure, that’s easily done with a prompt. But is that hole in the washer constrained to be concentric? Probably not.
Sure, LLMs are great at coding. Programming code is, after all, a language. And what does LLMs stand for? Large language models. What are they not so good at? Shapes, sketch constraints, feature dependencies, geometric relationships… the stuff of design intent.
The shape may be there. The “design intelligence” isn’t.
“What the market wants is for AI to understand its domain expertise, its design intent, and make a robust model that it is editable” in a reliable way. Current text-to-CAD approaches, he argued, fall well short.
The FeatureScript Advantage
PTC’s answer is FeatureScript — a CAD-specific programming language built into Onshape. It is exclusive to Onshape. FeatureScript was originally intended to let users make their features — hence the name. But it can also be used to create parts and assemblies. We see wave springs with FeatureScript as an example.
FeatureScript is essentially a English-like language for CAD. It turns out be very useful for AI.
When we created FeatureScript, we had no idea how useful it would be for AI, says Henry. LLMs had not been invented yet.
“We wanted to give users the ability to code their own CAD features.”
But users found that it can create geometry, it understands the fundamentals of shape. It an query faces and edges, understands constraints and relationships. Changes? Humans can edit the result within their ordinary interface.
A Token Effort
Companies are beginning to worry about tokens, the currency of AI, and Henry addressed them. When you use a traditional prompt-to-geometry approach, you’re spending tokens every single time the model generates or edits geometry. That gets expensive fast, especially when iterating toward a working design.
With the FeatureScript approach, tokens are spent once — to write the code. After that, the feature runs natively in Onshape. No additional tokens needed.
“AI created the tool, and we’re benefiting from it,” Henry said. “It’s an extremely economical approach to using AI tokens when building engineering content.”
PTC isn’t doing text-to-CAD. As his mentor John Hirschtick puts it, they’re doing “text to text to CAD” — text to code, code to CAD.
From Washers to Heat Exchangers
Henry walked through a progression of AI-generated FeatureScript examples, each more complex than the last. A roller chain sprocket was built by feeding Claude a web page defining the tooth profile and a manufacturer’s catalog page. The result: about 400 lines of code, a single encapsulated feature, correct tooth geometry.
We have transitioned all the way from a washer to a gear sprocket cluster. It’s a big jump but convincing.
Then the Heat Exchanger.
Henry gave credit his colleague Michael LaFleche, who set out to create a shell-and-tube heat exchanger using Claude version 4.7 and the Onshape FeatureScript MCP. LaFleche spent about a day creating the prompt, an extreme example of prompt engineering. It paid off.
The exchanger includes ASME and ASTM standard flanges with bolt-hole orientations automatically optimized to avoid interference with inlets — an unexpected and pleasant surprise. Baffles. Selectable head styles: flat, rounded, domed, or hemispherical. Weld prep cuts. Hot and cold inlet/outlet indicators (red and blue for hot and cold were another pleasant surprise. A fluid visualization followed that showed hot water in red and cold water in blue. A second feature that calculates weld time by going to the web, finding arc time data for each material thickness, and estimating accordingly. A cut list for all interior tubes. And self-generated documentation written into the design notes.
Final tally: approximately 2,100 lines of code. About 532 configurable parameters. Twenty total prompts — Henry thought it could have been done in fifteen. When you make a change, it rebuilds in 19 seconds.
“I believe that this is the most complex AI-generated model on the planet.”
He got his applause.
One Day of Prompting, 150 Hours Saved
A whole day spent on prompting at first seems silly. But Henry’s estimates the exercise saved 150 to 200 engineering hours — work that would have required, in his words, a rare combination of skills: CAD expertise, mechanical engineering, thermodynamics, and welding knowledge.
“It’s amazing,” Henry said. “You wouldn’t find that in one engineer.”
AI, in this case, does not replace engineers (the elephant in every room in which AI is discussed), but it compresses interdisciplinary work that would otherwise require a team, a long timeline, or a unicorn hire.
The Onshape FeatureScript MCP is currently in what PTC calls “labs” — available under a labs label in the near future, Henry said, as the Onshape Labs group continues developing it. A similar CAD-specific language is in development for Creo.
The Bigger Picture
Henry’s framework for AI in CAD is Advise, Assist, Automate.
Onshape’s AI Advisor is RAG based. RAG is Retrieval-Augmented Generation. It’s a technique for making LLMs more accurate and current by giving them access to a specific body of documents at query time, rather than relying solely on what they learned during training.
Onshape now fields more queries than all other help sources combined, by 30%. Creo 13’s AI Assistant offers contextual awareness of the model and what’s selected. Automation. We see Bobcat engine fuel-line optimization as proof.
Clearly, the heat exchanger is the statement piece. It’s PTC’s answer to the sea of startups that are doing text-to-CAD AI. And the FeatureScript abstraction layer is genuinely interesting.
Henry’s point is well taken: getting an LLM to understand sketch constraints well enough to build a sprocket tooth — 12 to 18 geometric primitives, roughly 80 constraints.
FeatureScript is clearly a differentiator, as much as, if not more than, Onshape’s millions of models created with its free version that AI can be trained on.