Anthropic previews MHS, a standard for AI agents to operate machines

Anthropic announced the Model Hardware Standard on Thursday, an interface that lets AI agents operate physical devices such as microscopes, liquid handlers and robot arms. The standard, MHS for short, is opening as a research preview with a first group of scientific research labs and advanced manufacturers. Anthropic plans to open source it after that.

MHS works through a standardized driver, software that translates between a computer’s operating system and a hardware device. The driver uses simple primitives, commands like read and write that any machine can act on, and it makes each device discoverable in a standard format. That lets devices and agents find each other across a network without a bespoke translator program in between. The result, Anthropic says, is that integration work that typically takes weeks or months drops to hours or minutes.

The driver also carries natural language tags for what code alone can’t show, such as the weight of a robot arm or the safety limits a device enforces. Those tags compile into a reference file that gives an agent everything it needs to operate hardware it has never seen before.

CNBC describes it as working like a USB-C cord, standardizing how information moves between devices. MHS is also model agnostic, which means users aren’t locked into Claude, and agents can control devices through the Model Context Protocol, a command line interface or code files. Anthropic open sourced MCP, a standard that connects agents to data sources, back in 2024.

We built this for science to sort of show the promise of AI, but there’s also huge benefits here for enterprise and for industry.

Elizabeth Kelly, head of beneficial deployments at Anthropic, via CNBC

The preview group includes AWS, Hugging Face, Raspberry Pi, Automata and Universal Robots, according to Ars Technica. Three research partners shared early results in Anthropic’s announcement.

Partner	Early result with MHS
Genentech	Automated a BCA protein assay across a liquid handler, a robotic arm and a plate reader
Carnegie Mellon	Ran dose-response experiments roughly 3x faster, with drivers and orchestration built in about 8 hours
UW Baker and Pinglay labs	Connected six instruments in under a week, including an agent-supervised qPCR run

Wired frames MHS as a set of rules for how AI agents should, and should not, interact with hardware. The caution isn’t abstract, because Anthropic and OpenAI have both found agents that hacked into outside systems and tried to deceive users during cybersecurity tasks, Wired reports. Anthropic says the preview partners help build safety evaluations and best practices before the standard becomes generally available.

The launch deepens Anthropic’s push into hardware, where rivals including OpenAI and Amazon have spent billions on AI native devices and manufacturing tools, per CNBC. The company is building a silicon team to design custom chips and recently hired hardware executive Caitlin Kalinowski, who previously worked at OpenAI, Meta and Apple. That push also includes Anthropic’s $45 billion Nscale compute rental, and it lands amid heavy investor appetite for robotics, like Generalist’s $198 million raise.

The thing to watch is whether MHS repeats the path of MCP, which went from company project to popular open standard. Alek Kemeny, the quantum physicist who co-led development, told Wired the impetus was “wanting to accelerate science.” He put the ambition in bigger terms in Anthropic’s launch video.

If you can test hypotheses faster, you could create general technologies faster. This is how a century of progress can condense into a decade.

Alek Kemeny, Anthropic, via Ars Technica

That depends on the open source release, and Anthropic hasn’t given a date for it. Until then, MHS stays a research preview, and the labs applying for access are the ones who’ll show whether hours-not-months holds up outside Anthropic’s own partner list.