Anthropic is making a move I didn’t see coming: out of the screen and into the lab. The new Model Hardware Standard – MHS for short – is a shared specification that lets an AI agent operate physical devices. Microscopes, liquid handlers, robotic arms. And several of them at once.
Setting all that up is usually a pain. Every device speaks its own language, few of them talk to each other, and anyone who wants to wire it all together spends weeks or months building custom integrations. MHS cuts that work down to hours or minutes. The trick is a standardized driver that translates between the operating system and the device, using simple commands like “read” (get temperature, say) and “write” (set temperature). On top of that come tags where you describe, in plain language, what a device can do and where its safety limits are. The weight of the robotic arm, for instance. That’s exactly what an agent needs to move it without breaking anything.
The whole thing builds on the Model Context Protocol that Anthropic introduced in 2024. And it’s model-agnostic: MHS works with any device that has a programmable interface, and with any agent, not just Claude. It grew out of a collaboration with the HHMI Janelia Research Campus, and Anthropic plans to open-source it once the preview ends.
The example that stuck with me most is from Genentech. Researchers there hooked MHS up to a BCA protein assay – a standard procedure that has to coordinate a liquid handler, a robotic arm and a plate reader. Claude was asked to optimize the flow rate for two very different liquids: plain water and a thick, foamy protein solution. It ran the transfers itself, read the plate data, and worked its way toward a human expert’s numbers. For water it landed around 140 µL/s, for the protein solution around 10 µL/s – values Genentech’s automation experts confirmed were reasonable.
What I appreciate is that Anthropic also shows the limits. When bubbles formed during mixing, Claude’s first instinct was to retry in the same spot with different parameters. That only made the bubbles worse, because it didn’t grasp the physics behind them. Only after the researchers explained that it had to move to a clean well and mix less did it fix the error – and keep the trick for the rest of the run. An agent that fiddles with a laser first and then writes a fixed script so the rest runs at the press of a button: that’s pretty close to how a person would approach it.
Physical AI was the chapter I kept waving off – too error-prone, too bespoke. MHS is the first thing that makes it feel like a real tool. For now it’s a preview for select labs and manufacturers, with partners like AWS, Danaher, Hugging Face and Raspberry Pi. If you want in, you can join the waitlist.
Sources: Previewing the Model Hardware Standard (Anthropic) · modelhardwarestandard.com