Anthropic opens a research preview of the Model Hardware Standard

The Model Hardware Standard comes out of Anthropic’s Beneficial Deployments team and a rig at HHMI Janelia, where a postdoctoral scientist had put an entire microscope setup into a shared memory dictionary so its instruments could talk at memory speed. The problem it addresses is mundane: most devices do not speak to each other, so a facility pays a specialist to write bespoke integrations, work the announcement puts at weeks or months.

MHS answers with a standardised driver and a small set of primitives, read and write, that any programmable device can act on, plus a discovery format so devices and agents find each other across a network. Tags let an operator describe a machine in plain language, including facts no code carries, such as the weight of a robot arm. The driver turns those tags into a reference file naming what the device measures, what can be adjusted and which safety limits are enforced. An agent reaches the hardware through Model Context Protocol, a command line, or code files it writes for long runs.

The numbers

QuEra Computing pointed Claude at a titanium sapphire laser whose frequency lock its engineers had already tried to automate. That bespoke script, built over several months, recovered the lock about 58% of the time at roughly 150 seconds an attempt. Claude ran unattended overnight as separate instances proposing changes, editing the script, running it against the live laser and reading the logs, and rewrote the linear sequence as a decision tree. Tested afterwards with no agent involved, the script recovered the correct lock in 695 of 700 trials, and the hardest cases took 10 to 14 seconds against the 5 to 10 minutes a person needs at the bench. Tuning the servo parameters separately brought a specialist’s standing 15.7 mV residual error down to 1.55 mV. At Carnegie Mellon, drivers written from scratch for a liquid handler, a plate reader, a robotic arm and cameras spread over three computers produced a finished dose response curve in eight hours.

What still needs a person

Anthropic is direct about the edges. Claude learns the physical world from text and images, so Genentech researchers had to tell it that a run failure came from foam in a well rather than from software, after which it held that context for the rest of the run. QuEra found Claude could not troubleshoot hardware faults it had no programmatic view of, and that it paused overnight waiting for approval on actions it judged risky. Devices with no programming interface remain out of scope, and access is by waitlist ahead of an open source release.