We’ve hosted multiple lab automation hacks in SF that have featured Physical MCPs that talk to instruments and robots, like the Opentrons Flex.
Many in the community follow up asking:
To answer those questions, let’s see what we can build with MCP’s + Skills + some hardware. 
Awesome, might have to check this out, won’t be able to make it in person but curious to see what comes out of it.
I’ve been using agentic frameworks (opencode and the million other clones) to automate my own workflows (in security rather than lab automation), so it’s great to see similar ideas being applied to this niche.
One thing I’ve been mulling over: it’s easy to over-engineer around current model limits (context windows, tool use quirks), only for those constraints to shift a few months later.
There’s also the practical side of MCP servers loading a bunch of tool definitions into context even when you’re only using a few, all that schema overhead eats into your working context. I’ve found it pays to keep the MCP/skills layer as thin and capability-driven as possible so it can adapt as models improve.
Part of me wonders if the endgame isn’t all this MCP plumbing and agent loop orchestration at all, maybe models just get good enough that you hand them a massive, well-structured system prompt and they execute reliably without needing the scaffolding. Less “tool ecosystem,” more “really good instructions.”
But who knows?
those constraints to shift a few months later
Yes, that’s just the nature of technology. I’ve co-hosted four AI × Lab Automation hackathons, and each one has been drastically different from the last. Some folks have attended all four and I imagine part of the fun for them is watching how fast things evolve.
On MCPs specifically, we held an MCP workshop on June 14th to test the idea of MCP’s with hardware and followed that with an official MCP hackathon on June 21st. We also just recently wrapped up an AI Science Cell Culture hack that used AI scientist tools & MCP’s to run experiments on a mystery cell culture. So we went from “let’s just see if it works” to “let’s run this workcell.” Here’s a vid if you’re curious.
But who knows?
Exactly. Our workshops and hacks offer a way for folks to “know” a little more as they get access to cutting edge hardware and software tools. It gives folks an opportunity to catch up if they’re swamped with work, school and life the other days of the year. We preface every event with “just a reminder that this is the worst these tools will ever be.” We also invite experts who can offer guidance and jump in on projects.
All of this is still nascent so lets just hack a little and find out.
If I weren’t on the other side of the country, I’d absolutely be there. As someone who started out in BME and gradually drifted into software engineering and security research, this intersection of agentic systems and lab hardware feels like the exact crossroads I’ve been unknowingly heading toward since graduation.
Very cool, just watched it! You might enjoy checking out Matt Brown on YouTube when you get a chance. He’s primarily a hardware hacker, but lately he’s been experimenting with autonomous frameworks built around Claude skills and MCPs… seems right up your alley.
https://www.youtube.com/watch?v=vMVU0UdAFMY
I’ve been working on something adjacent, a framework that can autonomously “drive” debugging and reverse engineering of firmware (specifically STM32F429xx devices) by giving the system access to SWD through a J-Link debugger.
Hack a little? Let’s hack a lot!