Anthropic Opens Research Preview of Model Hardware Standard for AI-Operated Physical Devices

The problem it addresses is plumbing. A lab bench or factory cell is assembled from instruments made by vendors that never planned to interoperate, so every device ships its own programming interface and specialists hand-write bespoke translators between each pair. According to Anthropic, such setup work normally takes weeks to months, and MHS reduces it to hours or minutes.

MHS standardizes the driver layer between the operating system and a device, exposing a small primitive set: read, write and discovery, so devices and agents can find each other across a network without a translator. Drivers can carry knowledge that code alone does not encode, such as the weight of a robot arm, using natural-language tags compiled into a reference file containing what a device measures, what can be adjusted and which safety limits are enforced. Control can be exercised through the Model Context Protocol, a command-line interface and code files. Anthropic said MHS is model-agnostic, meaning any agent harness can reach it via standard protocols.

The report detailed customer testing. Genentech automated a BCA protein assay across a liquid handler, a robotic arm and a plate reader. Claude, Anthropic's AI model, ran trial transfers of dyed liquid, read absorbance and scored itself against an expert's plate using root mean square error, converging on about 140 microliters per second for water and 10 microliters per second for viscous BSA. Genentech's automation experts confirmed the parameters as reasonable.

QuEra Computing reported the most striking gains. A bespoke laser-relock script built over months by a four-person team worked about 58% of the time, at roughly 150 seconds per attempt. Given the same problem through MHS, a four-role agent loop ran unattended overnight and produced a deterministic Python script that recovered the lock 695 times out of 700, or 99.3%, with the hardest cases solved in 10 to 14 seconds against 5 to 10 minutes for a human. Claude also cut the servo's residual error from a specialist's 15.7 millivolts to 1.55 millivolts; over a 19-hour run, its tune never lost lock, while the expert tune unlocked about 1.6 times per hour, according to a QuEra blog.

Carnegie Mellon ran dose-response experiments about three times faster by orchestrating a liquid handler, plate reader, robotic arm and cameras across three computers with incompatible interfaces, one of which had no programmatic interface at all. From driver-writing to a completed curve, including an autonomous rerun after the agent rejected a fit with R-squared below 0.9, took about eight hours compared with several weeks for a vendor setup. Six induced fault conditions were blocked before any device moved. At the University of Washington, a PhD student connected six instruments in under a week, driver-writing included. Tetsuwan Scientific paired MHS with its ResearchOS platform for qPCR pollution profiling. At Janelia, one microscopy rig went from seven programs launched in a fixed order to a single dashboard click.

Anthropic said MHS remains gated and still requires supervision. The report noted that Claude's physical reasoning has real gaps. Safety limits are stored in the driver rather than in the prompt, and the standard is compatible with the Model Context Protocol.