Agentic Work... Embracing MCP, Agentic Harnesses... To Survive

I’ve been watching a wave of AI-driven clones of key CAD tools hit the market—and specifically, the open-source community. What blew me away wasn’t just the clones themselves, but the use of agentic tooling and harnesses to create workflows that actually bypass the weaknesses of the underlying tools.

I love SketchUp. I’ve gone toe-to-toe with the architectural community—who have swallowed the Revit pill whole—to keep our small company on SketchUp. It’s my favorite tool for modeling factories, which is what I do. (Though I still rely on Fusion as part of my workflow, so this isn’t an anti-Autodesk rant.)

Personally, I think Trimble hit the sweet spot with their licensing—it makes the financial side easy.

But here’s the point of this post: Trimble needs to embrace the open-source community and the power of agentic tooling and harnesses. They shouldn’t settle for baking in a generic AI chat sidecar like everyone else is doing (and usually doing poorly).

For the past two weeks, I’ve been driving SketchUp directly from the command line. I’m writing product specs and building .skp components that are agentically placed into larger models. I’m taking two systems I’ve developed over time—a production line modeler and an ERP-lite purchasing engine—and connecting them to SketchUp to drive drawings and client proposals.

Doing this without a native Model Context Protocol (MCP) server has been an exercise in scaffolding. I built a file listener to monitor Ruby (.rb) scripts and leveraged native macOS automation to capture viewport snapshots for vision model feedback. I’m still early in the process, but the possibilities are huge.

Early on, I ran DeepSeek v4.1 and hit a lot of churn. I temporarily switched to OpenAI—burning $20 in under 10 minutes—which cracked the bridge implementation. Once I fed that architectural pattern back to DeepSeek, execution turned into a night-and-day difference.

I’m curious if anyone else here is tinkering with this kind of setup. If you’re building custom bridges, rigging file watchers, or hacking together agentic pipelines for CAD, I’d love to hear how you’re tackling it.

Trimble has a clear opening to get ahead of the open-source clones here. If they open up to native agentic integration—and honestly, considering how much 2027 LayOut actually rocks (anyone who knows my usual posts here knows I usually spend my time trashing LayOut)—it would put some real muscle behind the tool we’re already using.

Cheers!

Still waiting for “slope” piping’ the lack of MEP is a big factor that why many go to Revit in the mechanical and electrical side of things’

I had to make my own add-on to temporarily deal with this, there is one add-on they are working on it’ hopefully this will come out soon.

Hi Michael,

Interesting setup. A file watcher on .rb scripts plus viewport snapshots for vision feedback is basically the loop that most of the community MCP bridges are built around, so you may be able to skip some of the scaffolding.

A few community projects already do this. One is sketchup-mcp2, a Python MCP server paired with a Ruby extension that runs a TCP server inside SketchUp and executes commands against the live model. It also has a viewport screenshot tool, which needs SketchUp 2026 or newer. It was forked from the original mhyrr/sketchup-mcp. Supex is another. It’s aimed at people with agentic coding experience, lets the AI write and run Ruby scripts in SketchUp, and keeps the scripts in a git-versioned project. I haven’t compared them, and I haven’t checked how well they run on macOS, so they may not fit your setup. Even then, they’re worth a look for the screenshot-feedback pattern and for how they handle SketchUp’s single-threaded Ruby. pypi.org

One caution if you go that route. These bridges typically expose an eval_ruby tool that runs arbitrary Ruby inside SketchUp, so keep the server bound to localhost. Binding it to the whole network with no authentication would be risky. pypi

I agree with your main point. A generic chat sidecar is much less useful than a proper programmatic interface that agents can drive. As far as I know, the community is currently ahead of Trimble on this, and an official MCP server or a well-documented agent-friendly API would help a lot.

Mech3dcad’s point about MEP and sloped piping fits the same picture. Gaps like that are where custom extensions and agent workflows tend to fill in.

Would be glad to hear how your production line modeler and purchasing engine integration turns out.