I have not tried this yet, but looks interesting …
I’ve started exploring another MCP-based AI agent for KiCad: circuitron
There are other MCP-based tools mentioned on this forum: KiCad MCP: IPC-API Based AI Integration for PCB Design and KiCAD MCP Server
It’s interesting that there have already been a fair number of attempts at this given that KiCad is a niche application in a niche area. It speaks to the breadth of applications to which AI is being applied: it’s like water that manages to seep into every available crack.
It also shows the importance of open-source tools like KiCad that serve as a foundation you can build upon (even more so if it builds out its APIs) that other people can then freely use and improve.
If this evolves far enough, we may finally get an autorouter that actually works.
Can anyone offer an actual example of a pcb design performed or augmented with AI? Even just a little example, from conception to a feasible board?
I must admit I was skeptical about the utility of AI to my work flow, and, as an old fart who started pcb design at 4x size on a light table pasting donuts and tape, well, you know how stubborn old farts can be
In the last month I dug into the AI process for a pretty complex pyqtgraph signal analysis program using chatgpt/copilot/claude/perplexity and was pretty surprised at how quickly I got (from scratch) some initial code working, and how I could add features and chase down error messages. I also saw how easily it can add perfectly proper looking code that is nonsense, but all-in-all I got several thousand lines of code running way faster than if I had read a bunch of tutorials and dug through APIs and such. I paid for a subscription for copilot to run inside vscode for project-wide scope and between that and some simple open browser tabs with the others it has all improved my productivity. Even simple things like adding type hints and plumping up docstrings is quite fast. I am a vibe-coding convert.
Anyway, I am still a bit skeptical about AI augmenting pcb design, but am likely in a don’t-know-what-you-are-missing stage. I have read up a bit, but a little online example would be interesting and answer a bunch of questions.
My son, who is a computer programmer and has an IQ of WOW, says for him AI is a decent auto complete at best. Maybe we are just pushing it too far, too fast because we are more lazy than inquisitive? Slow down and let it evolve naturally instead of trying to push it too fast? Still garbage in, garbage out?
It’s not there yet for PCB design, but the writing is on the wall.
It could perform at the same level on PCBs as it does on code, but the money has not been poured into that application due to the small market size.
Just like in coding, the tech has not reached the level that it will replace 30 year veterans, but it will massively elevate hobbyists/entrepreneurs.
Hey Hermit: I was skeptical about AI for coding until I got started. Copilot in vscode improves auto-complete from the basic level, but it also has project-wide scope, so I can request things like “change docstrings from sphynx to google format in all files” and it highlights proposed changes in red/green for individual approval or rejection. Or add widgets to the gui layout, or add this method to compute… Yes, it needs to be carefully monitored, but I am coding much faster.
Here is an example you can copy/paste right now into chatgpt (free not even logged-in version) and get working code in a couple of minutes:
pyqtgraph/pyside6 program to open 4 channel 500sps 24bit wav file,
plot raw data with an adjustable epoch region,
plot time-domain epoch, plot fft of epoch,
and plot each channel in different colors.
Also plot scaleogram of channel-1 epoch using cmor4.0-4.0
and plasma colormap.
Add checkboxes to enable/disable channel plotting.
For epoch plots use mouse wheel to zoom y-axis, right-drag to zoom x-axis, and left-drag to pan x-axis.
Now paste the generated code into a .py file (and add the needed packages to the venv), run it and bada-bing, bada-boom you get:
Frickin’ cool imho. It needs lots of tweaking, but what a motivating way to get a project started.
I just have no idea of the current state of pcb AI. I doubt I would use AI to do system-design or even draw schematics, since I am pretty anal about the way I want my schematics to look, but there are possibilities for board layout, especially if nets can be defined as noisy buses to bury, fast nets to keep short and apply pair rules…
My son does banking software. He is responsible for every line submitted. I know him and his capabilities. He has had a keen interest in AI for over a decade or two and has followed it with keen interest. He even told me about algorithms he wanted to try out at least 15 years ago when AI was still a pipe dream.
Even as a kid he would get a new video game and sit down and read the manual/instructions completely before ever inserting the distribution media. So forgive me but you aren’t going to move me much here.
I took a little coding in college. My company paid for it so I went back to school and was one of those irritating older people in classes. I was taking things like C++ and machine language while others were gushing over things, I think in VB, that let them code spiffy looking programs that gave them a sense of awe with what ‘they’ created when it was mostly and literally window dressing provided by the widgets.
Bottom line? He IS using AI and I believe from past conversations has had access to all of the M$ stuff. He said their tools were really hard to beat. So, if he tells me he only trusts if for auto complete, that informs my opinion.
Yes indeed – I also need to tweak and refine every line as coding progresses, and would never blindly accept an ai-generated block of code as production-ready. My point is how easy it is to get a project started and incrementally add features without the tedium of poring through API docs. The little example generated from my prompt would have taken hours to get running, but I got a starting point in minutes and can progress from there. It eliminates some drudgery.