Anthropic has opened a research preview of the Model Hardware Standard, a specification that lets AI agents drive lab and manufacturing hardware: robotic arms, liquid handlers, microscopes, and other equipment with a programmable interface. The company is limiting access for now to a first group of research labs and manufacturers rather than releasing the standard broadly.
The stakes go beyond convenience. An agent that mishandles a software task can be rolled back with a git revert. An agent that mishandles a robotic arm or a liquid handler can break equipment, ruin a sample, or injure someone standing nearby. That is a materially different risk surface than the chat and coding agents Anthropic has shipped so far, and the company is treating it that way: it says it is building a physical safety roadmap and using the preview period to develop safety evaluations before considering an open-source release.
MHS works through what Anthropic calls a standardized driver, software that sits between a device and the agent controlling it. The driver exposes simple commands, “read” and “write” style instructions such as checking a temperature or setting one, and describes each device in a common format so agents can discover and address it without a bespoke integration. That reference file also carries context a specification sheet would not, such as the physical weight of a robotic arm, information Anthropic says has traditionally lived in paper manuals or in an engineer’s head.
Anthropic frames MHS as model-agnostic: any agent framework can reach a device through standard protocols, including the Model Context Protocol, Anthropic’s own tool-calling framework. That framing matters commercially as well as technically. A one-lab standard that other labs’ agents can plug into is a stronger position than a Claude-only integration, and it lowers the switching cost for hardware vendors deciding whose ecosystem to build for first.
The results Anthropic is publicizing come from its own launch partners, which include Genentech, Carnegie Mellon, the University of Washington, HHMI Janelia Research Campus, and QuEra Computing. QuEra reported that an agent running its laser-stabilization system recovered a precise laser lock without human help 99.3% of the time. Carnegie Mellon says a dose-response experiment ran roughly three times faster than before. Those figures come from Anthropic’s own writeup and its partners, not from an independent evaluation, and none of the participating labs have published the work through peer review.
Anthropic is also candid about where the system still struggles. At Genentech, researchers had to teach Claude to distinguish a physical failure, foaming in a protein sample, from a software bug, because the model’s understanding of the physical world comes from text and images rather than direct sensing. Anthropic describes this as a limitation that still requires expert oversight, which undercuts any read of MHS as a hands-off system today.
A roster of hardware and software vendors is building support into their own products: Amazon Web Services through its Strands Robots library, Tecan for its liquid handling platforms, Universal Robots for its robotic arms, and Hugging Face for its LeRobot library, among others named in Anthropic’s announcement. None of that amounts to a finished standard. MHS does not yet work with devices lacking a programmable interface, and Anthropic has not said when, or whether, it will formally open-source the specification.
For labs and manufacturers evaluating automation vendors over the next few months, the relevant question is not whether MHS works in Anthropic’s demos but who controls the standard once it leaves research preview. A specification for connecting physical equipment to AI agents that stays inside one company’s ecosystem is a lock-in mechanism; one that genuinely opens up is closer to a public utility. Anthropic says it intends the latter. Whether it delivers that before a rival lab ships a competing spec is the thing to watch.
Published by Anthropic on 28 August 2026.
