Anthropic is pitching its first move into physical AI as an open bridge between intelligent software and stubbornly incompatible machines: a faster route to autonomous labs and factories without forcing users into Claude’s ecosystem.

The company opened a research preview of its Model Hardware Standard, or MHS, on Thursday, beginning with scientific labs and advanced manufacturers. The premise is straightforward: equipment ranging from microscopes and liquid handlers to robotic arms generally speaks its own proprietary language, leaving facilities to spend weeks or months building custom integrations.¹

MHS seeks to replace that patchwork with a common driver and a small set of commands such as “read” and “write.” Anthropic says the system makes devices discoverable, gives agents operating and safety information, and lets them coordinate tasks across multiple instruments. In its account, the payoff is not merely convenience: agents could monitor results, adjust parameters and recover from some hardware errors while experiments run around the clock.¹

That ambition is being tested rather than declared won. Anthropic says early partners include Genentech, Carnegie Mellon, QuEra and Janelia, where pilots ranged from protein assays to laser calibration. At Carnegie Mellon, the company reported serial-dilution experiments ran about three times faster; at QuEra, an agent restored a laser’s required frequency “99.3% of the time without human intervention.”¹

The commercial argument is equally pointed. Elizabeth Kelly, Anthropic’s head of beneficial deployments, said the standard was built for science but carries “huge benefits” for enterprise and industry.² Yet MHS is explicitly model-agnostic, designed to work with Claude, rival models and open-source systems. That openness, Anthropic’s Jonah Cool said, is meant to counter lab hardware’s “proprietary solutions that are very brittle” and avoid vendor lock-in for scientists.³

For now, access remains limited and the standard is still a research preview. Anthropic says it will use the early rollout to develop safety evaluations and operating practices before open-sourcing MHS—a crucial qualification when the AI agent is no longer just generating text, but moving a robot arm or steering an experiment.¹