Anthropic is introducing the Model Hardware Standard (MHS) as a research preview, and it gives AI agents a direct line to real-world machinery. Microscopes, liquid handlers, robotic arms, even the laser systems inside quantum computers can all plug into the same interface.
Getting equipment from different vendors to talk to each other usually means building custom integrations — a job that can take weeks or even months. Anthropic says MHS can shrink that to hours, sometimes minutes.
Think of MHS as a universal hardware layer. Once a device is running an MHS driver, it tells agents what it can do, what state it's in, and what safety limits it won't budge on — all in a standardized format. Agents then control the machine through MCP, the command line, or an API, without hand-writing a new integration for every piece of gear. MHS isn't tied to Claude either; other models and agents are welcome.
The testing has already moved onto real machines. Carnegie Mellon University stood up an experimental automation system that would normally take weeks to assemble in about eight hours — and sped up its experiments by roughly 3x. Quantum computing company QuEra, meanwhile, had Claude automatically optimize a laser recovery routine; it succeeded 695 times out of 700 blind trials, a 99.3% success rate. AWS, Hugging Face, Raspberry Pi, and Universal Robots are also getting on board with MHS.
But MHS isn't a public standard yet — it's only open to a select group of research institutions and vendors, and Anthropic plans to open-source it only after more safety testing. Claude's understanding of the physical world is still limited, too. During testing at Genentech, it couldn't tell that liquid bubbling was a physical malfunction; researchers had to remind it this wasn't a software problem.