OnShape Labs Features AI Tools

Waiting months for “the next big CAD release” feels out of step with how fast AI is evolving, so we sit down with Darren Henry (SVP, PTC) and Cody Armstrong (Senior Director, Onshape AI Innovation) to explain what Onshape Labs is really designed to do. It is a broad, user-accessible early access program for what’s next in Onshape, where experimental features can ship faster than the normal three-week release cadence, get real feedback early, and then either graduate into the production cloud CAD product or get shut down quickly.

We also unpack the first Labs projects and why they matter to working engineers. The Omniverse Publisher, available via the Onshape App Store, focuses on robotics by moving CAD assemblies into NVIDIA Omniverse and Isaac Sim with far less friction. Because Onshape mates capture mechanical intent and degrees of freedom, teams can define joints and physical properties like stiffness and damping, then run kinematics and dynamics simulations in Isaac. The result is a tighter CAD-to-simulation loop for robotics design, physical AI workflows, and faster iteration between design changes and realistic motion behavior.

Then we go deep on FeatureScript MCP, which turns “custom features” into something far more accessible. Instead of relying on fragile text-to-CAD outputs, the approach is text-to-code-to-CAD: an LLM generates FeatureScript, inserts it, tests it, fixes errors, and iterates until it works. The best part for teams is that the LLM cost is mostly upfront, while the finished custom feature runs natively and can be shared across the organization without ongoing token usage. We cover demos, token realities, and why persistent memory could eventually capture company standards and tribal knowledge for the next generation of engineers.

Transcript

Welcome And Guest Introductions

Roopinder 0:07

Hello and welcome to FoDES, the Future of Design and Engineering Software Podcast. My name is Roopinder Tara. On the show, we will have guests that will discuss tools and technology that engineers will find interesting and useful. The purpose of our discussion is Onshape, the Onshape Labs, right? Recently announced. Do you want to just uh introduce yourself, Darren? Just say who you are?

Darren Henry 0:31

Hi, Roopoinder. My name is Darren Henry. For your listeners who don't know me, I've been in CAD a really long time, probably a veteran of CAD. I worked at SolidWorks for years. I'm now the uh senior vice president of general operations at PTC, leading a number of groups. We really focus, my groups really focus on growth and customer experience. It's everything from marketing to customer success and technical services.

Roopinder 0:55

I think I remember from PTC NEXT Darren, you said uh your whole life was CAD and you really have no other passions. No other interests? Yeah, no other interests.

Darren Henry 1:07

That may be true. I have a couple dogs that I uh spend a lot of time with, but yeah, it's CAD and dogs. CAD and dogs. Maybe food.

Roopinder 1:17

Oh, maybe food, okay. Right. All right. And hopefully, beer, aren't you talking? Weren't we talking about getting together on the West Coast for one of those brew breweries? I think the CAD Hardware Group is or the hardware group is doing another one soon.

Darren Henry 1:34

Yeah, no, we're actually, I just came out of a meeting where we were planning some West Coast events. So uh yeah, I'm sure we'll be back out there.

Roopinder 1:41

Okay. All right, very good. Cody, do you want to give a quick introduction to yourself for the podcast?

Cody Armstrong 1:47

Sure. Uh hello, everyone. My name is Cody Armstrong. I'm the senior director of Onshape AI Innovation, and I lead uh the a lot of the AI initiatives here at Onshape from a product management perspective. I've been in the cat industry now uh 20 years and with Onshape since the earliest startup days. So pre-acquisition with PTC. I'm really excited to dive into it.

Roopinder 2:10

Yeah, thanks for being on the show and uh great to have you. And Cody, you're you're modesty, you're one of the best demo jockeys I've ever seen. And I've seen a few. It's very kind, thank you. Know the product. Probably uh I hope you're not as well no now that Darren's not so one-dimensional, but I think your life centers around CAD too.

Cody Armstrong 2:32

It does. It does.

Roopinder 2:34

Good, good.

What Onshape Labs Actually Is

Roopinder 2:35

All right. So we're gonna talk about the labs that just recently announced on Shaped Labs.

Darren Henry 2:40

Yeah, I mean, this is this is huge for us. Um Cody, why don't you why don't you introduce the labs concept and talk a little bit about the projects that we have?

Cody Armstrong 2:48

I do have a slide that we can dive through that that kind of describes what labs is. Um so I'm gonna start with just a simple explanation as to what is on shape labs. And I think that's gonna be the most common question that we get asked. And in short, it's really an early access program for what's next in Onshape. So we want to be able to iterate even faster than Onshape has iterated in the fat in the past, right? And so you know we have a three-week cadence, so we release really quickly. But the idea behind labs is we want uh a place for us to release even faster and get feedback even faster than the traditional means that we have in the past. Um and so the idea behind Onshape Labs is it gives a broad amount of users, right? The idea behind labs is it'll be available to everyone, uh, gives a broad amount of availability to early access programs, including AI, before broadly available and brought into Onshape in the quote unquote production environment. Um and so it gives users, uh, a much larger pool of users than we've ever had in the past through traditional like EVP programs, uh, the ability to explore these new features and give us feedback. And we get feedback from a much larger pool uh of people and we can shape it earlier before it hits again, you know, the the on-shape proper. And so it allows us to also quickly experiment with different tools. And one of the things we found with AI is that there's just so many places where it can be used and where it can add value. And so the question then becomes where do we focus our energy? What is maximum value for our for our users? And so this gives us that that testing ground to experiment with the latest in AI capabilities and figure out quickly does this work or does it not? And an important part of the philosophy for labs is it is either adopted into on-chain because it sees success, or it's abandoned and we stop using it. Right? It's it's there's nothing that should persist in labs forever. Um, and it's just really a qu a vetting ground, a proving ground, so to speak, for AI capabilities and others as well.

Roopinder 4:44

I'm imagining a lot of AI agents now that are gonna be here eventually and they're gonna help us help me with all my engineering tasks, a lot of them. Is this gonna be all developed by PTC or is it all gonna be is it are you going to have like an app store, like the Apple App Store?

Cody Armstrong 5:03

So is the question will third parties be able to build into it? Yeah, yeah. So we will enable third parties to build on top of labs features as well. Absolutely. Oh, okay. Yeah, but Onshape already has the app store, right? And we already have actually a number of different AI focused apps in the app store, um, which is really exciting to see. Um, but yeah, that we want to be an open platform for sure.

Roopinder 5:25

Okay. All right. This does not replace the app store. This is an addition to this is where your this is your lab, just your lab. So this is where products are getting developed. Got it. Okay.

Darren Henry 5:36

Yeah, and in fact, Roopinder, it's really interesting. The different projects we're doing will be deployed through different mechanisms, and some will actually be deployed right in our app store as you can think of them as little add-ins. Some will be deployed as features within the software that you'll enable, and some will be external to Onshape that you'll connect through various means. So it's really exciting, you know, the wide variety of projects that we're thinking about. And each of them will be deployed in an easy-to-use manner uh based off what the project is.

Roopinder 6:11

What's the intention then to kind of like get a uh like trials up there so you can judge feedback or or you know, I'm used to a world where you didn't announce anything before it was ready to ship. Now we're entering a world where this stuff in development is being made available, right? So yeah, to tell me more about that sentiment, like what's happening here.

Darren Henry 6:38

You know, it's really interesting. If you're not familiar with Onshape, we put out releases every three weeks. You know, traditionally CAD companies had sort of annual user conferences, they'd announce what's coming next year, and then you'd wait like six to nine months, and then new functionality would roll out. We're actually on a different cadence, and we put out releases in our core product every three weeks. Well, when it comes to these like these new technologies, everything from AI to, you know, we're talking about an Omniverse publisher, these are technologies where we can move even faster. And what's nice about this is we'll have concepts that we think will be usable by our community. We now can go independent of our standard three weeks cadence, put them out there, get feedback before we refine it. So it's an earlier cycle where the customer can actually help define the ultimate state of the feature, or tell us it's of no value, and then we'll actually either graduate it in product or we'll terminate the project. Uh, but what's really nice is what's the motivation? The motivation is with AI and these new technologies, the world's moving very fast. We have a talented team that can put out concepts very quickly, we get feedback and we decide which way to go. So that's really the the the driving force behind labs. It's a good idea.

Roopinder 8:01

I mean, I'm seeing I'm seeing so many. I'm seeing not a day goes by, I don't see uh uh something that's gonna help me. It's using AI for for CAD, for either for design or for engineering. It's uh there's so much out there. And uh so yeah, I'm certainly welcoming of this early look.

Darren Henry 8:21

And and I'll I'll say one more thing is like the the name labs, we went back and forth on it because so many software companies have had labs over the years, but we, you know, it's not just isolated to us. The lab-like projects, these experimental concepts will be, I think, are you're just gonna see a surge across all industries. There's a lot of things that we built as a foundation to on shape that have immense benefit now in the world of AI. And um, we're very fortunate to have not only a deep project set that we're working on, but a deep well of ideas for future projects as well.

Roopinder 9:00

Are these ideas coming internally or are they coming? Sorry, Cody, we'll let you talk soon.

Cody Armstrong 9:05

Both sources, I would say. Um I would say both sources, uh a lot of internal inspiration. We have some examples of that, but also very much customer driven.

Roopinder 9:13

Okay. Any any journalists giving you ideas? No, probably not, right? Not so much, not yet, but we're we're always open, Retender.

Darren Henry 9:21

Cody, maybe we should talk a little bit about the first project that we released.

Shipping Faster Than The Release Cadence

Darren Henry 9:25

Do you want to talk about that?

Cody Armstrong 9:26

We want to talk about feature script MCP, which is our first project.

Darren Henry 9:30

I was thinking more about the Omniverse Publisher, the one that's now available.

Cody Armstrong 9:34

Let me bring that up. Is available in the Onshape app store. It's really intended for robot, at least our implementation of it. I think that there's lots uh that you can do with it, but our implementation is really focused around robotics and incorporating robotics into Isaac Sim, which is the NVIDIA simulation environment.

Darren Henry 9:52

Um, and so yeah, there's two there's two applications in the NVIDIA suite, Isaac Sim and Isaac Lab, where you can actually do you can actually do analysis, sort of dynamics analysis, kinematic analysis on your robots. And then Isaac Lab actually allows you to train, virtually train and put your robot through different scenarios virtually, and then take that information and use it to train a physical robot. So it's exciting that CAD data can easily get into this environment.

Cody Armstrong 10:25

Yeah, and I think that's the nuance that I point out is easily get from a CAD model to a functioning simulation. And robots has traditionally been pretty difficult. Um, and and Onshape's architecture and its mates in particular make this well suited. Um, but obviously, our ability to ship features like this really quickly means uh we can you know allow a company who's building robotic simulation to use Onshape to build that and export it out and build their training runs uh inside of it.

Roopinder 10:51

Well glad to see it because it was years ago I heard Jensen Long introduce uh physical AI, and part of that was Isaac. And uh and then it could, you know, and then I heard, oh, it's gonna make it's gonna make humanoid robots possible because it's going to help you simulate their motion. And I thought, oh, it's great, but it's some not until now I've actually seen anybody use it.

Darren Henry 11:13

Yeah, which this is one of those sort of technologies that it was one of the underpinnings of Onshape, a decision in Onshape's early days that made a connection to Omniverse so robust today. And Cody alluded to it. You know, we define our mates, our assembly mates a little bit different. We use sort of a point-to-point system with degrees of freedom. What we're really doing is capturing mechanical intent. And then with this publishing application, you can actually define physical properties around those joints. So a robot would have could have dozens of these joints. You're defining the physical property right in Onshape, and then you're passing that information the mechanical intent of the actual joint and the physical properties of the joint, you're passing it to Omniverse. Omniverse reads those automatically, so there's no setup, and it can start running simulations. If you then learn from the simulation that you need to make a design change, you make the design change in Onshape, you fetch the data into Omniverse, and it's just a clear pipeline. And what's really cool about this is Jensen Wong on stage at GTC, GTC is NVIDIA's big event, they actually showed running the simulation, AI watching the video of the simulation, making a recommendation of the change in Onshape, changing it, and then rerunning the simulation for a success. So it's this is one that could be utilized and has a tremendous value outside of AI, but if you can couple AI to it, it can build a pipeline where you can iterate very quickly between systems.

Roopinder 12:52

Yeah, nice loop going on there and all in real time. It's uh it's just uh it's mind-boggling. Uh so the idea is that now the robot can handle. You mentioned the dynamics, the kinematics of the robot, uh, which takes into account the physics, the mass, the forces. Uh, it can it now now CAD with Omniverse and Isaac will have a concept of that, right? You will have it won't just be your mates like pieces snapping together. Parts will actually understand contact and gravity and not be able to.

Darren Henry 13:28

Yeah, and and what you're what you're really defining is the stiffness and dampness on each joint. And you're you're sort of allowing the the arm of the robot to be rigid, but you're defining the properties, the physical properties of the joints. And then as it's going through motion, it's doing that dynamic analysis. So you can start understanding, you know, with gravity, will it be able to pick something up? When it grabs it, is there a large vibration? Those types of things, or large, you know, uh large motion, large large shift due to the motion of the robot. So it's it's yeah, these are these are things that you can now simulate and uh and use that in a nice way to inform your next design change.

Roopinder 14:05

Now we could always do that, right? I mean, not always, but since we had computers, uh we could always do that with atoms, right? Atoms uh are uh any of those kinematics. But now what does tell me, just tell me what what uh what makes Omniverse? I think it's yeah, do that, do that, does it do it sort of automatically?

Darren Henry 14:27

I think it's two things. I think it's the performance of the system, and I, you know, and the other thing that's really nice is the Omniverse libraries, where they have maybe as a CAD person, I'm building an end effector of a robot and I'm using an off-the-shelf FANUC robot. They have a wide variety of libraries available and then different environments and different settings that you can add as well. But but it's also the performance of the system. It's a GPU-based system. You can build very large factory floors and simulate that in a in a very realistic way. Oh, okay, okay, got it. So this one, this one, and I I don't want to jump on your your uh your host, Cody, but this one is available today. So our customers can go to the app store, they can subscribe to this app. You can see it's a free and you can hit the subscription button, you can plug it in, and now you'll be able to see this interface appear on the right side of Onshape and publish out to the file formats needed uh for Omniverse. Okay.

Cody Armstrong 15:27

And this was released just last week. So this is as we announced last public.

Roopinder 15:31

So take my assembly, whether it whether it be a robot or four-bar linkage or anything, it would just take it into Omniverse and it would behave like it would in real life. Is that correct? Yeah, it is.

Cody Armstrong 15:42

You define the mates as you would in Onshape. And one of the beauties of this is it uses Onshape's mating system to define those joints, as Darren mentioned. All right. The process from going from there to joints that are defined is very easy.

Roopinder 15:55

So I don't have to learn atoms or use it or anything like Ansys or just it's all there. It's all no, it's all free. It's free. How long are you planning on keeping that up?

Cody Armstrong 16:04

I don't think we have any plans to change that in the short term. Yeah, I don't think so. I don't think I don't see anything. And it's important to point out this still requires you to go to Isaac Sim and run the simulations there and things like that. This is a tool that helps facilitate that exchange. Okay. All right, so we want to dive into the next topic.

Omniverse Publisher For Robot Simulation

Darren Henry 16:19

Yeah, let's talk about MCP.

Cody Armstrong 16:20

All right. So there's actually a whole host of things that we're focused on for labs that you're going to see soon. And I think you've seen some demonstrations of them in the past, but I really wanted to focus, and Darren's really excited about this one as well. Um, and that is FeatureScript MCP. Um we already have this in available in the labs program. Um, so there's already users using this. And we're really excited about the implications for it for the future. Um, so what is it? It's a purpose-built MCP server for FeatureScript. So if you're familiar with UnShape, we have the FeatureScript language, which is the language we use to generate geometry. This is an MCP server that's purpose-built to generate FeatureScript code. Right. And so the the idea here is we want to make FeatureScript, a language that generally requires programming knowledge, available to everyone. And so the average user, even though they have no programming knowledge, can still generate really useful features that saves their whole organization tremendous amounts of time. Now it's not just generating FeatureScript. So it will generate FeatureScript and it will not only generate it, but then insert and test it and evaluate it on its own independently. And that's what really makes it a unique thing. It's a loop that runs largely autonomously, right? So you give it a goal and it goes through and generates code, it tests the code, it evaluates that output, and then it makes a modification and does that loop again until it sees that goal as being a bat. And so it's really exciting from an end-user perspective because you can very easily connect to the on-shape FeatureScript MCP, give it a simple prompt, maybe some supporting materials that support what you want to build and let it run. And let it run in the background. And it will connect to Onshape and communicate with Onshape and build the necessary pieces uh for that particular custom feature. So it not only creates the custom feature, but it also evaluates, tests and evaluates it as well. So it'll figure out on its own what errors were generated in that uh FeatureScript, fix them, and continue on. Um and this is available in Onshape Labs. And we really do think this uh democratizes custom features to everyone. Right. In the past it's been very, I don't want to say niche, but it's been specific to those that can program. And now the door is wide open. And the the example of this is just internally in the last few weeks we've had an explosion of new custom features in every conceivable way. Dozens, this is only a handful, but dozens of new features created internally in just the last few weeks. Um and so this is a hurdle that's gone from taking days to weeks to hours to minutes, right? And it has big implications. Like some of these examples you see here are specific industries like camshafts or acme threaded rods, where that may be an operation that takes 20 minutes, now it takes 20 seconds, and I can distribute it to the whole team. And that's really the value in FeatureScript MCP is we can take this modeling operation that took you hours, turn it into a custom feature that takes you just a few clicks. And anyone can do it.

Roopinder 19:21

That's the encrypting languages, you know, there every CAD program has them. Uh, would you say that this is going to really make feature more usable or use more user-friendly now than ever? Or just make more people use FeatureScript? You know, I'd probably, even though FeatureScript is powerful, would you say the minority of on-shaped users actually take advantage of it? And now it's going to change?

Cody Armstrong 19:57

That's the goal. Yeah, because we believe FeatureScript is a real powerful differentiator. And we've talked to customers who've completely reworked their workflow with FeatureScript and save a tremendous amount of time as a result.

Roopinder 20:08

So this lowers the barrier to using FeatureScript. I I could be an engineer's like a frame may not be the right term, but hesitant to go into programming. And now I can with FeatureScript MCP, I can I can go there and generate even looks like we're fairly complicated FeatureScript.

Cody Armstrong 20:26

Correct. Yeah. And that's that's the idea behind it, is we really want to open up that audience to anyone. And the result, it's important to point out, and Darren is going to stressing this, is the resulting feature does not require an LLM. So the LLM is only used to generate the feature, and then you can share that custom feature with everyone in your organization, and there's no cost, uh no LLM cost associated with running it, right? So it's a one-time cost associated with generating that feature. Um, so I think that will be especially important for a lot of users out there. Is there's not this persistent LLM cost just to generate new geometry. Once the feature is generated, it's done, and you can share it just.

Roopinder 21:02

So users are LLMs initially, but then from then on it's using FeatureScript. Okay. Correct. Got it. So we save those tokens. Save those tokens. Exactly.

Darren Henry 21:12

You wanna you wanna see a demo of it, Roopinder? Yes, please.

FeatureScript MCP Changes Custom Features

Darren Henry 21:15

All right, I'm gonna stop sharing.

Roopinder 21:17

I'm gonna try to share. Yeah, I don't know if you want to take it if you're taking requests, but I would maybe this is your demo. And we did not talk before this. Are you gonna make a gear?

Darren Henry 21:28

Well, I'm gonna show you a couple things. Do you guys see my screen?

Roopinder 21:32

Yeah, yeah. Sorry, we can see.

Darren Henry 21:35

So the first thing I I wanted to show you is is what the interface would look like. Cody had it up on this the image, but what how I run it, and you can run it a lot of different ways, but how I run it is I run Claude in a console. This is Claude. It can mute work with any LLM, but I run it in a console. And here's what I wrote I wrote new project, let's make a custom feature in Onshape that is a flange pushing, that is a Flange bushing standard sizes. I didn't even use proper English. And what Claude will do is say, great, new project, a flange bushing generator, love it. Let me frame it before I touch the API. It then says, which standard? You could do metric or you could do ISO, and it says, or SAE, sorry, metric or SAE. And I say, let's go, look, let's do both. Let's do both. Let's give me a selection. And then it actually says, okay, great, I'll fetch the information. What's key is I'll stop. I stop this one here. It says, Do you want to allow Claude to fetch this content? So it found some information, in this case, an IGIS um website. And then it prompts you. So it's always, you know, you're always in control. It prompts you. Do you want me to get this info? Is this what you're looking for? And there is some back and forth as you work. Some examples, and this is one that I talked about on stage at uh PTC Next that you might have seen. Was uh this is a sprocket where I asked it, hey, let's create a sprocket. I know a lot about sprockets because I did a tutorial on it years ago. But a sprocket has really interesting tooth profile that requires a lot of construction geometry.

Roopinder 23:12

Yeah, this is your uh what's that curve? Invalid curve, invalid spot.

Darren Henry 23:17

This one's not an involute, but it is a lot of circles with tangencies. So you're thinking of a um of a gear tooth, but on a sprocket, this is a roller chain sprocket. You do have sort of the complexities. As you can see, if you look, you can see we have a lot of different surfaces that govern each tooth. So I asked it to make sure it had the tooth profile correctly. I pointed it to a web resource. I then pointed it to a web resource of a sprocket manufacturer to get standard sizes, and it built this custom feature. The key is complex geometry. You see, fillets, champers, the tooth salves taper, editable, where we can now edit this. This is what the feature looks like. Hopefully, you can see this on your screen. Where I can do things like I can change the sizes. It's grabbing the Martin Sprocket catalog. So I can change from a 12-tooth to an 18-tooth to I'll go back to a 12-tooth. I can decide if I want a single strand or a double strand. Maybe it, you know, two different chains will run off the sprocket, maybe three different chains. So look at the complexity of the geometry. That's number one. Number two is look how editable is. Do I want set screws where it puts a set screw hole at two at 90 degrees or one? Or I can do forget the set screw. I want a keyway. There's the keyway. Maybe I want to customize this parameter. Maybe I just want to buy the blank and I take off the center hole altogether. Or maybe I want a little bit longer hub, maybe half-inch hub. And the key is I don't want the hub on both sides. I want the hub on one side. And this is all AI driven. It's a toolkit it but it built me. It runs native and as Cody said, it used tokens when it built a tool, but as I'm using it right now for you, I'm not using tokens. And the key is it captures that design intent. We've seen a lot of examples of people building things like sprockets, but you can't edit them without going back to the LLM. Here, we've built a tool that you can utilize throughout your whole company. By the way, the beauty of Onshape, you hit the share button. I can share it with you and Cody. If you have Onshape, you can use a Sprocket immediately. So it really helps you drive fast automation throughout your company.

Roopinder 25:39

The second one I want to show you, I should show the code, Darren, before you just want to ask you about that one because I wouldn't say trip me up as a right phrase, but I've used a gear for people that have said have uh claimed that they understood geometry as we know it, as engineers know it. And uh, you know, they fall pretty flat. And most of the gears they don't mash or take chains or belts or anything. They they're they're obviously not gears. So my question here is does it recognize each of those tooth, each of those teeth as a tooth, or is it still like a CAD geometry?

Darren Henry 26:20

So it it's it's still CAD geometry, but you can do you can ask it to calculate different different aspects of it. Like, for example, if I wanted to calculate the weight of this and change the material, I could just prompt the LLM to add a different material, pull down for me to do stainless steel or brass or whatever I need, and it will calculate certain properties. It it in this case, it's not seeing the tooth. The AI recognizes this as a tooth. If I have a problem, I can say there's a problem with the tooth taper, and it actually recognizes that the tooth taper goes in two directions, and it will say radial tooth direction or or longitudinal, I guess, and it will help me prompt it. So it understands that, but the geometry itself doesn't have any more intelligence than what a human would build.

Roopinder 27:09

Oh, okay. All right, okay.

Darren Henry 27:10

Uh, Cody brought up a good point, which was the main point I missed, which is this is the magic. We didn't do text to CAD, we did text to code to CAD, and this is what the LLM was able to generate using the Onshape Labs FeatureScript MCP. So it's an expert at programming the language behind Onshape. And in this case, it generated over 700 lines of code, and it even documented it really well for me so I can understand it. If I did know how to program this, I could go in and make edits manually. But the beauty is I could easily just go back in and ask Claude to make edits to it if I wanted another feature, if I wanted to change the color of that sprocket or have a pull down for another different size, I could do that. So that it's it's very easy. Plus, one last thing, because Onshape has data management built in, as Claude's making changes, I can easily version control the new features. So I could have version one of the Claude, Sprocket Builder, version two, version three, version four, as I go. And I can even ask it to document all my prompts for me and store that in on shape as well. It's quite powerful.

Roopinder 28:32

I'm looking at it, and I'm I'm not a coder, but I even I can uh appreciate how well that code looks structured, is even comments in it. That's just uh wow.

Cody Armstrong 28:42

And the beauty of this, just to reinforce this, is he can share that with anyone in the organization, and now they have a sprocket feature, right? That they don't have to do anything to do, and there's no LLM cost associated with running it in the future.

Roopinder 28:53

So is that a FeatureScript feature now? That what I'm looking at, that that code length of 700 lines of code, that's a FeatureScript.

Darren Henry 29:02

Yeah, so so the language is called FeatureScript, and what it wrote is a custom feature that generates a sprocket. Okay, all right, and it's really key that there's a lot of different things you can do with FeatureScript. So, for example, this is an assembly, it's a it's a wire shelf assembly. This has multiple parts. You can see all the different parts here. Um but what I'm what I did here was I I said, hey, I'd like to be able to select. Sorry, let me uh show you this. I'd like to be able to select a rectangle and a circle. And if I edit that, I can edit it to any distance. Let's make this 20 by, you probably have a big sink at home Roopinder. I'm gonna put 31.

Roopinder 29:53

Okay.

Darren Henry 29:54

And we'll sorry, I'm I do actually and we'll make the drain a little bigger, we'll make it 5.5. Okay, and just for kicks, I'll just move that drain over here and I'll accept that. And when I roll forward, this was a a custom feature that I made again through prompts. I actually showed it a picture of what I was trying to do, and I'm like, I want to make a wire frame. I want you to give me the ability to change the thickness of the wire around the frame.

Roopinder 30:26

I want to make it you're doing this live, right? This is not those wires just popped up there. Oh, yeah.

Darren Henry 30:31

This is live. And by the way, everything we do at Onshape is usually live. The code doesn't crash. This is we demo live all the time. This is live with AI driven made code that I don't even know how to program, and it's doing it. 1.5 spacing, and I'll make this one uh 2.5. Okay, and you can see how it's changing it. And the beauty is look at this, Roopinder. You like real time? I'll edit this, I'll hit final, and I'll just move this guy over here and it recalculates everything for me. Nice, nice.

Roopinder 31:09

And if if I don't like the spacing in some places, I could I could change that, but it's a good, it's a good first, definitely a good first start at it, right?

Darren Henry 31:18

So yeah, and and you got to think somewhere there's an engineer that has to lay out wire shelves, wire racks. He can prompt it with Claude, build a tool, pass it with his entire company, and now they're able to do it very quickly. I would guess to get the cut list and to do this correctly, you're talking hours. It's probably an hour's worth of prompting, and now you have a tool that will save you hours every time you have to build this. Yeah, it's amazing. Um, another great thing that I wanted to show you is this.

Live Demos Sprockets And Missing Tools

Darren Henry 31:49

There are features that were missing in Onshape that you might like, exotic features and other CAT systems. One of them is actually, you know, if I go to this corner here and I hit chamfer, um, we can pick edges when we chamfer, but we can't pick a vertice. So here I can pick an edge and I can pick an edge. But I may want or a surface, but what I may want to do is actually just chamfer, not the corner off. And some CAD tools have that, and you might miss that. Maybe that's one you wanted. 18 minutes with Claude. I asked it to build it, I gave it a picture, asked it to build it, and now I have this feature called corner chamfer, where I can pick any vertex and real time it's doing it. And even if it's a complex surface, no worries. This is just grabbing it, it's allowing you to knock that corner off, knock that corner off, knock that corner off. So the days of like feature comparison of CAD wars is over because if it doesn't exist with us, you just type into Claude or type into Chat GPT or or Copilot, and it'll program it for you with FeatureScript MCP. We have a lot of different examples of this kind of thing.

Roopinder 33:05

It's really I gotta ask you, are you you're I'm sure you're not using the free version of Claude, right? This is the higher end, is this the that's a great question.

Darren Henry 33:14

So it will work with the free version, but what happens is you'll run into your limits. So uh a lot of the work that I've done, a lot of these features I built with the $17 a month Claude plan. But full disclosure, you know, we have an enterprise version where I've probably spent a couple hundred dollars on different features. And yeah, tokens are expensive part of this. And I think it's always going to be. I think you know, most of the engineering community, there's going to be individuals at the company that do eat a lot of tokens per month. So you are you're kind of at a all-you can eat buffet of of tokens, but uh but the ordinary person, how quickly do you think they would have exhausted their their free their limits with uh with just a yeah, I I don't think I you know, I think everything that I've done to date, including you know the sprocket, everything that showed you, I have a lot more examples. Yeah, I don't think I've spent a thousand dollars worth of tokens. Um so worst case, I think if you're very if you're very active with it, I don't think it's gonna cost you more than twenty thousand dollars a year on the most active. And I think the average will be let's say somewhere around twelve hundred to fifteen hundred dollars in tokens. Got it.

Cody Armstrong 34:32

But that would be for someone who's using it constantly, like it's constantly building new features for your business, right? So the ROI there would be quite large.

Roopinder 34:40

Maybe you're a program CAD manager and you're doing a lot of these developing these feature scripts for your group, you might be doing it all the time. In that case, I am I contracting with Claude for this, or is some of this getting through uh coming in from just from using Onshape? Do you do that?

Darren Henry 34:58

Yeah, so another great question. So so the the key is you need an LLM, whether it's Claude or you know, you could use GitHub Copilot or you could use um Chat GPT. So yeah, you're gonna need LLM tokens. So you're gonna contract with somebody on that. But there are also API limits with something like Onshape because the one thing that I'm not showing you is I could run several Claude sessions simultaneously with Onshape and do my own work, and I can consume a ton of API calls. So we do have additional API limits for our customers that are using it extremely uh heavily.

Roopinder 35:36

Okay.

Darren Henry 35:37

Okay, I wanted to show you two more examples, and um, this one I also alluded to the last time we met, which was um a fill feature. So I showed you like we can build parts and systems of parts like the wire shelf. In this example, I actually built a um a tool, not actually something that generates geometry of a product, but something that can actually help me do some engineering work. And in this case, what it is I have this little tank, and this tank actually was imported. But what I wanted to do was understand the volume of this tank. So I asked Claude to build me a tool which would fill this little tank with a volume of liquid that I specify. And here you can see it's telling me that uh it's it can hold 0.68 liters, but I filled it to 0.35 liters. We'll just go ahead and change that to 0.5 liters, and you can see it very quickly calculates. Now, what's really neat about this, this was V1 of my tool, and you can see it's limited in choices. In V2 of my tool, I just um I just picked a a sort of a an odd square sort of welded tank, maybe you'd find on a boat or something like that, a barge. And on this one, I said, you know, what else could we do? And it actually suggested that we do different types of liquids. So maybe I want salt water, or maybe I want honey, or I want motor oil. I have the ability to tilt the liquid, tilt in different directions, right? I'll turn off tilt and it'll be flat. I have the ability to specify the volume or the mass. So there's 15 gallons, and it also tells me the weight now of the liquid. It tells me the the capacity of the tank is 19 gallons, and it says 19 gallons because I have a hole here in the side. If I suppress that hole, you can see it goes up to 40 gallons, which is uh, you know, the full amount of the tank. So imagine I'm a medical device manufacturer and I have to do graduation marks on maybe a micro titer tray or something. I have now a tool that can fill up a volume and give me the ability to put marks, graduation marks, on any container.

Roopinder 38:05

Or I'm not sure if it's irregularly shaped or not constant process.

Darren Henry 38:10

Exactly. So it's really a powerful thing. Build the features you want, build the parts you need, build the systems of uh of components that you might like, or build engineering tools to help you.

Cody Armstrong 38:23

And one thing that I would add, it's a very unique aspect of FeatureScript MCP is the persistent memory. In fact, Aaron can show it here, but but what's happening each time you generate a new feature is it's remembering what you did in the last time and it's avoiding any issues and subsequent features. And so over time, your agent will get more and more in tune with what you do and avoid issues less and less.

Roopinder 38:43

So it's it's watching everything and it's learning.

Cody Armstrong 38:45

It's learning persistent memory, yeah. And so it's actually built into the toolkit. It writes a persistent memory down to a document inside of your uh Onshape library, and then it reads from that for every new request. So it avoids the same errors, it knows what you're trying to do, it knows what you've done in the past and the kinds of features that you commonly use and will get better and better

Tokens Costs And API Limits

Cody Armstrong 39:05

with time.

Roopinder 39:05

All right, this may be an annoying habit of journalists. You know how you tell them something that you did that's really cool, and then they ask for, oh yeah, well, can it do? They ask for more, right? So is there a way to apply that sort of individual learning for each user to the group, to the company?

Cody Armstrong 39:22

There's a few ways that that we could disseminate that information, and we have a handful of ideas around that. We actually did that a lot internally. So a lot of the default uh um skills that are associated with these toolkits are learned just from people like Darren and myself who've been using internally and kind of distilled down a common thread. What we want in the future is for these to be company specific and for you to be able to, as an organization, not just an individual user, but as an organization, learn and persist. And that just that applies more, that applies to FeatureScript , but it also applies to things like drawing standards and you know, release management standards and a whole bunch of other practices where you know each company has their own kind of standards.

Roopinder 40:02

Wouldn't that be great if I could uh you know, as a young engineer, I could uh I could draw upon the uh tribal knowledge, if you will, of the entire company when I'm starting to make something from scratch. I had this, I still remember traumatic, it was traumatic for me to have to go through and try to make something the first time and go up against senior engineers who were uh knew how to make it like the back of their hand. You know, they would do it, know it designed. Um I'm going through guidebooks and I'm not finding anything. It's honestly, it's such a great thing to have a collect this all amount to a collective knowledge, a collective knowledge, a growing collective knowledge for the company.

Cody Armstrong 40:41

Yeah, yeah, that would say very important priority for us.

Roopinder 40:44

Yeah, okay, great.

Cody Armstrong 40:45

So next time we meet, working on that. I think it, you know, you'll see it in rollout. So you FeatureScript MCP is available, and that then that will improve on it in just usual on-shape fashion every three weeks, right? Or even faster in the context of that.

Darren Henry 40:57

Yeah, one last thing I I want to say is those of you familiar with Claude know that there's Fable is the latest model. I use primarily Opus, and uh and we've done a lot in Sonnet, and Sonnet is the older model. What's also exciting about this isn't the efficient token usage that we have, or it's not only that, but it's also the fact that the models keep getting better and better and better, and you don't need the latest model to build some of this geometry. I mean, we're proving it right now with Opus and uh and Sonnet. So it's um every day it's things will be in the will will benefit the engineer. I think the models will get cheaper, I think the compute's gonna get faster, and with Onshape, you have a scalable system that can take advantage of both very quickly.

Roopinder 41:46

Yeah, I'm a cloud user too, and I know that Sonnet 5 that started doing a lot of the things that uh the higher end models were doing, and I could save some money, right?

Cody Armstrong 41:57

It'd save it and we think that trend will continue. Uh I believe Anthropic created the standard and it's been adopted by other people.

Roopinder 42:04

Okay, but the people that are using MCPs are using Claude as well.

Cody Armstrong 42:09

But there's nothing specific about it, so you can connect with Gemini or any any uh AI that supports the MCP protocol and most too.

Darren Henry 42:16

Did I end with one more example and then we'll call it a day?

Persistent Memory And Company Standards

Darren Henry 42:19

Please, yeah. All right, let me share. This one's for you, Roopinder. I think you'll appreciate this. Let's go back to my desktop. One of the one of the things I wanted to do, it's sort of a fun example, is you know, I was thinking about what's complex geometry that I wouldn't build by hand, but maybe AI could figure out. And what I came up with was a gemstone. So brilliant. Yeah. And what's really nice is look at the feature. Again, the code isn't that isn't that uh great. It's uh less than the sprocket, 300 lines of code that it built. But um, you know, this is just because I just think you're a gem, Ruby. You could do round, we could do oval. Uh, we could actually maybe give it a different color, a ruby.

Roopinder 43:09

I I think you you might have used what you did use before where you were just lopping the corner off something, but uh and applied it in in uh succession. But or did you just say do a brilliant cut on this shape?

Darren Henry 43:21

So this did take a lot of prompting. Um, it had to go out and research how to do it. But what was really interesting, this is a cushion cut diamond. Okay. What's really interesting here is to get the faceting correctly, I was uploading images and it would converge on it called it like the keel length of the facet. It was using terms that I didn't know, but it actually converged from images that I uploaded and got the result. So I just think I just think if anybody that knows CAD out there would realize this is would be a challenging, challenging part, even to figure out where to take advantage of. symmetry would be would be challenging. And again, I have one feature, I would say less than four hours of prompting, where if I was a jeweler and building jewelry, which is a great application for on shape, I now have my gemstones ready, ready to do that.

Roopinder 44:18

It might be replacing uh Rhino. Surprising enough, I think Rhino is using her gems for uh jewelry making. Yeah, we may have the AI advantage in that. Maybe you'll surprise me with some woodworking tools. That's my that's my that's your jam. Yeah that's my jam. All right well that's been great very informative and I'm very excited you guys are making good use of AI. AI has been absolutely my focus the last few I would say last at least the last year and uh I'm glad to see it being implemented so well. I'm very impatient with CAD companies dragging their feet back up and just refuse to say that LLMs have any use in CAD creation. And uh I'm so I'm so against that idea. CAD is a language after all right the language of CAD is is a limited language compared to like English right and LLMs do great with English. I don't know why it can't be used to make things right as a front end. I I get it CAD companies are great at creating geometry let them do that but the interface could should be can and should be LM style right okay but there you go I'm just ranting.

Darren Henry 45:29

No I I I think you're preaching to the choir here um we we we have that unique advantage in that we expose that FeatureScript language and and it's really a powerful advantage for us.

Cody Armstrong 45:40

And I would also follow it up we have a strong culture of delivering and we we believe strongly in and getting it in the hands of users. And that's why FeatureScript MCP is available today. You know as we we prioritize that and also I think you have mentioned this in the past that because Onshape is on the cloud it's using I'll say infinite compute resources that it can actually do a lot of that neural processing better than and even in a world you can multi-thread in a level that it would be impossible for a human and and you can imagine a world where you generate 20 drawings simultaneously just using AI, right? So the the potential on shape's potential in this is just uh it's incredible. And so you're gonna see us I think you're gonna see us take more and more advantage of the architectural benefits um that we've built a decade ago now in this new context of AI.

Roopinder 46:33

And another advantage I'd give to you and I think I'm sure you thought mentioned it before is that uh so many of the models are available for public use and and feeding the learning used for machine learning for for AI.

Cody Armstrong 46:48

That's uh most CAT companies don't don't have the uh means or the rights to use customer models right but you guys have this vast library of the public models yes yeah the public model space which is you know Franchi free users uh is available to everyone and there's tens of millions of parts in that space that we can learn from so absolutely and and some of the improvements you'll see soon are are related to that in search and other areas in the public space so stuff we're definitely excited about.

Roopinder 47:17

Excellent

Gemstone Demo And Closing Thoughts

Roopinder 47:18

very good all right gentlemen thank you so much unless there's anything else we could no I think we we covered we'll we'll have more in a few weeks but um we appreciate your time we always do all right great to see you and uh thanks Cody for the demo and uh great to hear what you guys again great to hear what you're doing thank you thank you talk soon bye everyone thank you for listening to FoDES the future of design and engineering software show brought to you by ENGtechnica. I hope you have learned of a new application or technology that will help you with your job. If you have an application you think would be of interest to us please let me know by emailing me at Roopinder engtechnica.com