On August 27, 2026, Anthropic announced the opening of the research preview of the Model Hardware Standard (MHS) to an initial group of scientific research laboratories and advanced manufacturers. The standard provides a shared specification that enables artificial intelligence agents to operate and coordinate multiple physical devices simultaneously, such as microscopes, liquid handlers, and robotic arms, to carry out tasks ranging from drug discovery experiments to laser calibration in a quantum computer.
MHS was developed through a collaboration between Anthropic and HHMI Janelia Research Campus. The company believes the core problem is that each device typically relies on a different software interface, forcing a laboratory or industrial facility to build custom integrations that can take weeks or months. The standard aims to reduce this work to hours or minutes, while enabling the agent to coordinate experimental steps, update parameters in real time, and, in some cases, handle device errors without direct intervention.
How does the standard work?
MHS adds a unified driver that translates between the operating system and the physical device. The software relies on a set of simple commands, such as read to read the temperature and write to adjust it, while also making each device discoverable in a standard format over the network. Instead of creating a custom translator for each device, the agent can find connected devices and communicate with them through the shared interface.
The driver also includes natural-language annotations describing properties that do not necessarily appear in the code, such as the robotic arm’s weight, safety limits, and what the device can measure or adjust. The standard converts this data into a reference file that explains to the agent how to operate the device. Equipment can then be controlled through the Model Context Protocol, a command-line interface, code files, and application programming interfaces, allowing multiple devices to be coordinated through a single line of code.
Early results and partners
Genentech used the standard to automate a BCA protein assay by coordinating a liquid handler, a robotic arm, and a plate reader. At the University of Washington, MHS was used to remotely monitor devices, oversee a qPCR reaction, and connect a robotic arm to a liquid handler to deliver plates without collisions. Researchers at Carnegie Mellon University ran dose-response curve experiments at nearly three times the previous speed, using devices distributed across three computers with incompatible interfaces.
At HHMI Janelia, the standard helped unify a microscopy platform that relied on seven software programs from different vendors. QuEra also used the standard to control part of the laser system inside its quantum computers, and the agent developed a mechanism for recovering laser lock with a 99.3% success rate without human intervention. Organizations also working to support MHS include Amazon Web Services through Strands Robots, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, and Universal Robots, as well as Hugging Face and Raspberry Pi.
What changes in practice?
If the specification matures and is released as open source as Anthropic plans, it could provide a shared layer for operating scientific and robotic equipment from different companies, rather than leaving each integration as a separate project. However, the announcement describes an early research phase, not an open standard that is already available to everyone. MHS also currently works only with devices that have a programmable software interface.
Anthropic acknowledges that Claude remains limited in spatial and physical reasoning and therefore requires expert supervision. In Genentech’s experiments, researchers had to guide the model to understand that foam in protein samples represented a physical failure rather than a software error. The company will use the preview to develop additional safety evaluations and a physical safety roadmap before releasing the standard as open source.