LabVIEW
LabVIEW MCP Released (Use Claude, GPT, Copilot… to create, analyze, and discuss your code)
After an extended beta phase, I’ve decided to release the LabVIEW MCP. At the start of the beta phase, I mentioned that many features were still missing before I could consider it a release.
The most important of these features are now functional, and we can use it in our daily work.
I’ve already used the LabVIEW MCP myself to create classes, unit tests, and libraries. The generated code isn’t always perfect, but it’s certainly usable.
That’s why I recommend that anyone looking to build a modern agent-based workflow use it to gain some initial experience. It gives us the opportunity to build workflows on a larger scale and move closer to a partially or fully automated workflow.
When it comes to workflows, I recommend taking a closer look at N8N and Claude Workflow. With AI, things move so fast that by tomorrow, other tools will already be the latest and greatest!
The MCP currently offers the following features (requires LV2026):
- Creating projects, classes and libraries
- VIs, Controls and SubVIs in the project scope
- User events (with some limitations)
- Class support
- Interfaces (Classes are not allowed as interfaces data)
- Typedef support (create, use)
- Creating controls (also enums)
- Projects, libraries,
- xnodes (with limitations!)
- Connector Pane in all orientations and patterns
- Create description and Icon (Sometimes the AI is lazy and forgets the icon 🙂 )
- Unit tests (Caraya, LUnit)
- DQMH (only create modules and events) => I am in contact with the Consortium to improve integration
- and some more agents for documentation and more...
Let me know where there are still issues. There's still a lot of room for improvement in terms of performance. I've implemented a few workarounds for the gRPC interface. As soon as AIXML is expanded, I'll be able to remove my workarounds, and we'll see improved performance.
As mentioned, we can of course do more than just generate code. We can also carry out code reviews, write code documentation, troubleshoot, debug and so on!
Here are a few examples of what the generated code might look like:
You can also find more information in the LabVIEW Discord channel/AI.
You can find the code here (All open source and MIT licensed):
https://github.com/Zuehlke/labview-mcp
https://github.com/Zuehlke/labview-mcp/releases
Currently, I recommend using the ZIP file from the Releases section (latest version). Unzip the files into a folder and then, for example, configure Claude according to the instructions.
=> We're still having an issue with our installation script, but it should be fixed soon.
In the future, I will implement an add-on/plug-in architecture in the MCP. The goal is for the MCP to replicate the core capabilities of LabVIEW, with extensions such as DQMH, Actor Framework, etc., added via add-ons and agents. Agents alone will not be sufficient, because that would consume too many tokens. Therefore, add-ons will be used to add additional tools to the MCP.
Have fun, and don’t forget to give it a star on GitHub 🙂
Claude (4D version)
TestStand users can also use TestStand MCP (fully featured):
https://github.com/Zuehlke/teststand-mcp
LabVIEW-MCP is now CLD certified (unofficially)
********************************************************
I ran the CLD Certification through LabVIEW-MCP. Everything worked the first time. You can see the results in the attachment.
The only thing I had to correct in a second command were the icons. This is a known issue the AI often forgets to include them. I still need to fix that.
You can also see that the comments are still a bit unstructured on the block diagram. This is another known issue that could be corrected with an additional command.
But I wanted to show you what it looks like today.
Procedure:
- Download the example and delete the solution! 🙂
- Point Claude or another AI tool to the directory
- Enter the following command:
In this directory, there is a PDF with a problem statement. Your task is to solve this problem using the LabVIEW MCP. Use the agents if appropriate, and read through the document first.
In the “Template” folder, there are VIs that you should use. - In the end, the code must be executable and fulfill the task requirements.
Finally, the MCP wrote its own little test and re-tested the specifications to make sure everything is correct.
I’m curious to see if I would have passed the CLD officially with this! 🙂
There are even more CLD sample tests available. Give them a go and see for yourselves what the current situation is!
Best
Claude
Last night, I instructed the LabVIEW MCP to run the ATM CLD test. It took exactly 52 minutes. However, 3 of those minutes were due to a LabVIEW crashes.
The test clearly showed that we’re still too slow, but the results are already usable. During that time, the code was built entirely based on the requirements and automatically tested at the end.
The GUI is still ugly, but it works.I haven’t trained the MCP on all the front-panel customizations yet.
=> Yes, I’m also going to teach the MCP not to implement ‘The VI’ and to pay more attention to the size! 🙂
My next goal is to reduce the build time from 52 minutes to 30 minutes.
Here are a few pictures from the ATM CLD exam:
@RTSLVU wrote:
@Cloedu72 wrote:
You can also find more information in the LabVIEW Discord channel/AI.
There's a LabVIEW Discord? I found it, I don't know why I never thought to look for one?
There's more than one.. List can be found here: https://labviewwiki.org/wiki/LabVIEW_Community_Managed_Discord_Servers
Have you tried it with any open-source/local only models? My company has a (small) local AI server. I've got the link running using Claude Code, but connected to this local server. So far I've treid qwen3-coder and Gemma4; qwen3-coder made a bunch of irrelevant classes, then couldn't make the real VI, and Gemma took about 35 minutes to just make a plan for a VI.
I'm guessing I need a better LLM model, but there are about a zillion of them out there and I have no idea which one would work best.
(We have some legal/IP restrictions that make cloud models, like Claude, a real problem).
