Claude Moves Into the Physical World With Anthropic’s MHS

Introduction

Anthropic is extending Claude’s tool-use model beyond software. The company has released a research preview of the Model Hardware Standard, or MHS, designed to give AI agents a consistent way to interact with physical equipment such as robotic arms, microscopes, cameras, and laboratory instruments.

The idea is similar in spirit to the Model Context Protocol, or MCP. MCP gives an agent a standard way to discover and call software tools, data sources, and services. MHS applies a comparable abstraction to hardware, where every manufacturer may expose different drivers, commands, data formats, and safety assumptions.

What the standard is meant to do

• Describe hardware in a common language. An MHS-connected device can expose its identity, current state, available actions, physical properties, and operating limits in a machine-readable format.
• Reduce integration work. Instead of writing custom glue code for every instrument, an agent can discover a device, inspect its capabilities, and call supported functions through MHS, MCP, a command line, or code.
• Support different forms of embodiment. Claude does not need to be permanently installed in a particular robot. A camera, arm, microscope, or other instrument could become part of a task when connected to the standard.
• Enable closed-loop interaction. Devices can return live results, allowing the agent to adjust parameters, respond to faults, and decide what to do next rather than simply issuing a fixed sequence of commands.

Anthropic says Claude has been used to control the low-cost SO-ARM101 arm in the Hugging Face LeRobot ecosystem. In the demonstration, the model measures the arm’s workspace, calibrates it, and then invokes the underlying controller. The setup does not rely on a pre-trained robot policy, teleoperation, or human demonstrations for the showcased task.

Why laboratories are an important test case

MHS originated in work with HHMI Janelia, where research systems often combine equipment from multiple vendors. A brain-imaging setup, for example, may include lasers, motorized focus components, and cameras, each with its own control software. Connecting them can require weeks or months of engineering.

One demonstration focuses on microscopy. When a moving cell leaves its original field of view, Claude can read the camera and microscope state, produce a tracking program and interface, and adjust the microscope to follow the target. That could reduce the amount of time researchers spend watching screens and manually turning controls.

The larger promise is not that an AI can press buttons, but that it can coordinate instruments while keeping their constraints visible. In physical environments, however, a bad action can cause a collision, contaminate a sample, damage equipment, or create an irreversible result. Descriptions of safety boundaries are useful, but they do not by themselves guarantee safe behavior.

MHS is currently available only as a limited research preview for selected institutions and hardware companies. Anthropic plans to use real-world testing to refine the standard, safety evaluations, and operating guidance before considering a broader open-source release. Its strategic significance is therefore still unproven. If the interface matures, the next layer of embodied AI may be less about building one machine for one model and more about making many kinds of machines reliably legible to an agent.

Source: QbitAI