Anthropic is developing a standard that allows AI agents to control a wide range of machines and research equipment. The Model Hardware Standard (MHS) is designed to eliminate the need to build a connection for each device individually. The technology is now available as a research preview.
MHS focuses, among other things, on equipment in laboratories, manufacturing environments, and robotics. Using the standard, AI agents can, for example, operate microscopes, fluid handling systems, robotic arms, and lasers. Anthropic initially developed the technology in collaboration with the HHMI Janelia Research Campus.
Standard layer between AI and machines
The approach is reminiscent of the Model Context Protocol (MCP), through which Anthropic previously introduced a standard to give AI models access to external data sources and software tools. MHS extends that principle to physical equipment.
To do this, MHS uses standardized drivers. These translate equipment functions into relatively simple commands, such as reading or adjusting a value. At the same time, each driver contains information about the device’s capabilities and limitations. This allows an AI agent to determine what a connected machine can do and the limits within which it may be used.
Ars Technica notes, however, that MHS is not dependent on AI. Devices can also be controlled directly via the command line or API code. The connection to an AI model is established, in part, through MCP. This allows an agent to process commands in natural language and determine which sequential actions are necessary.
This is primarily intended to reduce integration work. Devices from different manufacturers typically use their own interfaces, forcing organizations to build custom integrations. According to Anthropic, connecting and integrating hardware can therefore take weeks or even months. With MHS, this could be accomplished in hours or even minutes in some cases.
Agent controls the entire workflow
It’s not just about giving individual commands to a machine. An agent can combine different devices within a single process, monitor results, and adjust settings as circumstances change. For tasks that must be performed over a long period or very quickly, the agent can convert commands into code that is then executed autonomously.
Anthropic demonstrated this, among other things, by adjusting a laser. Claude changed the setting, used camera footage to assess the effect, and repeated the process. The model then converted the knowledge it had gained into a deterministic script that allowed the laser to be aligned without continuous intervention from the AI model.
In another demonstration, Claude got a robotic arm to pick up an aluminum can, even though the model had not been specifically trained for those actions. An agent can also, for example, focus a microscope, analyze the results, and then determine which area requires further examination.
MHS is not tied to Claude. The standard is model-agnostic and can be used via various interfaces, including MCP. Anthropic’s ultimate goal is to create a common hardware layer that various AI agents can connect to.
AWS and robot manufacturers are on board
Several technology and hardware companies are already experimenting with the standard. AWS is working on support via Strands Robots, while Doosan Robotics is testing MHS with robotic arms. QIAGEN is exploring applications for laboratory equipment, and Tecan is building support for its automated liquid-handling systems. Universal Robots also plans to integrate MHS into its robotics platform.
In addition, Hugging Face is working on support within the LeRobot robotics library. Raspberry Pi is also experimenting with MHS and has tested the technology with cameras, among other applications.
For now, MHS remains a research preview. Anthropic points out that AI models still have limitations in spatial and physical reasoning, which means human expertise remains necessary. During the testing phase, the company therefore also intends to further develop safety measures and evaluation methods.
After the preview, Anthropic plans to make MHS available as open source. This should ultimately allow the standard to be used outside the Claude ecosystem as well, for connecting AI agents to the physical world.