Anthropic opened a research preview of the Model Hardware Standard (MHS) on August 27, 2026, sharing a specification meant to let AI agents operate physical devices safely. The first group includes scientific labs and advanced manufacturers. MHS is designed so agents can run multiple instruments in parallel, from microscopes and liquid handlers to robotic arms, on tasks that range from routine drug-discovery assays to laser calibration on a quantum computer.

The work started as a collaboration between Anthropic and HHMI Janelia Research Campus. Anthropic says connecting a lab or factory floor usually takes weeks or months because devices do not share a common interface. MHS is meant to cut that integration to hours or minutes, then let agents sequence steps, update parameters in real time, and in some cases recover from hardware errors without a specialist standing over the rig.

Drivers, MCP, and early users

MHS introduces a standardized driver that translates simple read and write primitives so devices can be discovered on a network without a custom translator for each vendor. Agents can reach hardware through the Model Context Protocol, a command line, or code files. Anthropic says the standard is model-agnostic and works with any device that already has a programmable interface.

Early projects named in the announcement include Genentech automating a BCA protein assay, University of Washington labs supervising qPCR and plate handoffs, Carnegie Mellon running dose-response experiments about three times faster, Janelia unifying a seven-program microscopy rig, QuEra recovering a quantum laser lock 99.3% of the time without human intervention, and Tetsuwan Scientific profiling pollution with qPCR. Hardware and software partners testing or adding support include Amazon Web Services, Automata, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots, Hugging Face’s LeRobot library, and Raspberry Pi. Anthropic is also exploring work with Danaher.

Anthropic is not open-sourcing MHS yet. The company says it will use the preview to build physical safety evaluations, and that Claude still needs expert oversight because it can misread physical failures as software bugs. Devices without a programming interface are out of scope until vendors add drivers.

Decoded Take

The interesting bet is not that Claude can run a liquid handler. It is that labs will accept a shared driver layer the way software teams accepted USB or MCP, instead of paying for another year of bespoke integrations. If Hugging Face, Raspberry Pi, and instrument vendors actually ship MHS drivers, the standard becomes a distribution play for physical AI that is harder to copy than a single lab demo. The risk is the opposite: a preview that stays inside friendly partners while safety reviews slow an open-source release. Watch which named vendors ship a public driver, and whether Anthropic publishes the safety evaluations it promised when the spec goes open.