Anthropic is moving its AI ambitions off the screen and onto the lab bench, betting that a common interface can turn disconnected instruments into coordinated, round-the-clock systems. Its harder task may be persuading equipment makers to adopt the standard widely enough for that promise to matter.

On Thursday, the company opened a research preview of the Model Hardware Standard, or MHS, developed with HHMI Janelia Research Campus. The specification is intended to let AI agents safely operate programmable devices—from microscopes and liquid handlers to robotic arms—while allowing several instruments to work in parallel.

The problem is familiar to laboratories and factory floors: each machine tends to arrive with its own interface, forcing specialists to build custom links that can take weeks or months. Anthropic says MHS replaces much of that work with a standard driver and a small set of commands, including “read” and “write,” cutting setup to hours or minutes.¹

The company’s case rests on automation with guardrails, not merely faster machinery. Device descriptions can include natural-language details and enforced safety limits, while agents can monitor results, adjust parameters and, in some cases, recover from errors. Anthropic is sharing the early system with labs, robotics firms, electronics groups and manufacturers to develop safety evaluations before it is open-sourced.²

Early tests offer the bullish version of that future. Carnegie Mellon researchers reportedly ran serial-dilution experiments about three times faster; QuEra said an agent restored a quantum-laser lock 99.3% of the time without human intervention. In another demonstration, Claude adjusted a laser, assessed the result through a camera, then converted its learning into a repeatable script.

Yet the standard is not a magic adapter: only equipment with programmable interfaces can use it, and older devices may need upgrades. Anthropic’s differentiator is that MHS is model-agnostic, built atop its Model Context Protocol rather than reserved for Claude. As partnership chief Jonah Cool put it, scientific gear often “suffers from proprietary solutions that are very brittle,” and the goal is to “avoid vendor lock-in for scientists.”¹

That makes the preview a bid to set the plumbing for physical AI—not just sell another model. Whether it becomes the lab’s shared language now depends on the vendors whose machines must learn to speak it.