Anthropic has officially introduced a new Model Hardware Standard (MHS) designed to help AI agents interact with and control physical equipment used in laboratories, factories and research facilities. Taking to X, the company has announced a research preview of the standard to a select group of research organisations and hardware manufacturers, with plans to eventually make it open source. It was developed initially in collaboration with HHMI Janelia Research Campus and wants to address a long-standing problem in physical automation: getting different machines to communicate with each other and with AI systems without requiring custom integrations for every device.
Anthropic says integrating laboratory and manufacturing equipment can currently take weeks or even months because individual machines often rely on proprietary programming interfaces. MHS attempts to reduce that process to hours or minutes by providing a common software layer between AI agents and programming hardware.
The standard uses drivers that translate basic commands between a computer and a physical device. These include operations such as reading measurements or changing settings. Each connected machine can also be described using standardised information that tells an AI agent what the hardware can do and which operational limits it must follow.
An MHS driver can include details that may not be obvious from software alone, such as the physical characteristics and safety requirements of a robotic arm. This information can be added using natural-language descriptions and is then converted into a reference file that an AI agent can use when operating the equipment.
Once devices are connected, AI agents can control them through the Model Context Protocol (MCP), command-line tools, or APIs. This allows an agent to coordinate multiple machines, monitor their output and adjust parameters as an experiment progresses.
Anthropic says Claude has already been tested with MHS in scenarios where it interacted with physical equipment in an exploratory way. In one example, Claude adjusted a laser, observed the resulting beam movement through a camera and repeated the process before turning what it learned into a deterministic script for future use.
Anthropic has been working with organisations across biotechnology, robotics, quantum computing and manufacturing to test the standard. Amazon Web Services plans to support MHS through its Strands Robots library, while companies including Doosan Robotics, Universal Robots, Tecan and QIAGEN are exploring integrations with their hardware and laboratory systems.
The company says early testing has shown improvements in hardware integration, experimental iteration and real-time fault detection. However, Anthropic acknowledges that AI agents still have limitations when dealing with physical environments and will require expert oversight.