Anthropic previews MHS AI hardware control standard

In experiments involving liquids, even moving a single tube requires connecting a liquid-handling robot, robotic arm and measurement equipment using separate methods. Anthropic developed the common specification Model Hardware Standard (MHS) with the US-based HHMI Janelia Research Campus to overcome this obstacle, and began its research preview on 27 August 2026. It is initially being offered to scientific research institutions and a limited number of partners in advanced manufacturing.

MHS is intended not to increase the types of equipment, but to standardize the entry points for operating them. Standard drivers consist of basic commands such as “read,” which retrieves the temperature, and “write,” which changes a value. Characteristics that cannot be determined from code alone, such as a robotic arm’s weight, are described with natural-language tags, while measurable values, changeable values and safety limitations are automatically generated as reference files. AI agents can control equipment in three ways: through MCP (Model Context Protocol, a mechanism for connecting AI with external equipment), the command line and code files.

Previously, custom implementations for enabling equipment to communicate with one another took several weeks to several months. Anthropic says MHS can shorten that integration work to several hours to several minutes. At Genentech, the company automated a protein quantification method using a liquid-handling robot, robotic arm and plate reader. Claude autonomously adjusted the liquid-dispensing speed and recovered from a failed attempt to retrieve a tip, Anthropic says. At Carnegie Mellon University, a dose-response test involving serial dilution of liquids was run at about three times the previous speed, with the process from a non-automated state to obtaining results completed in 8 hours. The same work would have taken several weeks if outsourced to a vendor.

The specific change is the amount of time researchers spend connecting equipment from different manufacturers. AWS supports MHS through Strands Robots, a library connecting AI agents with physical devices, while Tecan and Universal Robots are also working on support and validation. Hugging Face plans to add MHS support to its robotics library LeRobot, and Raspberry Pi says it will add support across its own product range. As the number of compatible devices grows, researchers could spend less time preparing to begin experiments that combine equipment from multiple manufacturers.

However, this is not a full rollout of a finalized standard. Equipment without a programming interface is not currently supported, and the process of integrating drivers with manufacturers’ equipment is still underway. In Genentech’s experiments, researchers also had to teach Claude that an error caused by foaming was a physical failure rather than a software bug. Anthropic says expert supervision remains necessary and plans to open-source MHS after establishing safety evaluations and safeguards.