Decoded: What is Anthropic's MHS and why does it matter for Physical AI

Anthropic's Model Hardware Standard aims to give AI agents a common way to understand and operate programmable hardware, potentially simplifying integration across robotics, labs and manufacturing

Picture a factory floor with a robotic arm from one manufacturer, cameras from another, sensors from a third and machines from a fourth. Each of these can be programmed individually, but there is no guarantee they speak the same language. This has been seen before with the Internet of Things (IoT).

 

Now, add an AI agent that is supposed to look at what the camera sees, decide a part is defective and tell the robotic arm to pull it off the line. Before any of that can happen, someone has to physically wire these systems together, expose what each one can do to the software and make sure the commands being sent never exceed what the hardware can safely handle.

 

 

That integration work is what Anthropic's Model Hardware Standard (MHS) is trying to shrink.

 

Anthropic opened a research preview of MHS last week, describing it as a shared specification that lets AI agents operate physical devices such as microscopes, liquid handlers and robotic arms. Its early rollout is concentrated in scientific laboratories, but the company says the standard is meant for any device with a programmable interface, spanning robotics, electronics and manufacturing. 

What is MHS?

MHS introduces a standardised driver between a computer and a physical device.

 

Think of the driver as a software layer that translates between the way a machine works and a standard language that applications can use. Instead of every application having to learn the proprietary interface of every machine, the MHS driver exposes the hardware through a common set of basic operations.

 

The standard also tries to describe the machine itself.

 

A robotic arm, for example, has a certain weight and physical range. Those characteristics matter when deciding how it can be moved. Anthropic says information such as this has traditionally been buried in manuals, stored on a user's computer or simply existed as knowledge in the head of an engineer.

 

MHS uses tags to capture such information, including what a device can measure, what can be adjusted and which safety limits apply. These tags can then be used to generate a reference file that an AI agent can use to understand the device.

 

An AI agent does not necessarily need to have been specifically trained on a particular robotic arm or laboratory instrument. If the device has an MHS driver and a reference file describing its capabilities and constraints, the agent can obtain that information when it needs to use the machine.

Where MCP comes in

The Model Context Protocol (MCP), which Anthropic introduced in 2024, was designed to standardise how AI applications connect to external data and tools. Instead of developers building separate connectors between an AI application and every database, repository or software service, MCP provides a common protocol through which those systems can be exposed to an AI application. It has since become an important connectivity layer for AI services.

 

MHS applies a similar idea to physical hardware, but the two are not the same thing.

 

MHS describes and standardises the hardware interface. MCP provides one of the ways an AI agent can access that hardware.

 

Anthropic also says MHS is model-agnostic, meaning it is not designed specifically for Claude. Any agent harness capable of using the standard can access MHS devices.

The control loop

In an account of an experiment shared by Anthropic, Claude calibrated a laser by nudging it, checking the result through a camera, adjusting it again and repeating the process to work out the relationship between command and outcome. Once it had figured that out, it wrote a script that performed the whole alignment as a single command, rather than reasoning through every step each time.

Similarly, at the University of Washington, researchers used MHS for an AI-supervised quantitative polymerase chain reaction (qPCR) process and for coordinating a robotic arm and liquid handler during plate handoffs.

 

These are laboratory examples, but the underlying pattern is much broader. The agent observes. It decides what needs to happen. It sends instructions to one machine, receives the result and uses that information to determine the next action on another machine.

From laboratory automation to industrial automation

This is where MHS could become relevant beyond the environments Anthropic is currently showcasing.

 

A modern factory already contains many programmable machines. Robots move parts, cameras inspect them, sensors monitor temperatures and pressures, and production equipment performs individual manufacturing operations.

 

The problem is that these systems are often built as separate components, with software written specifically to make them work together.

 

A common hardware interface could make it easier to put an AI agent above those systems.

 

For example, an agent could receive a high-level instruction to inspect a batch, coordinate cameras and sensors, instruct a robot to move suspect components and adjust a machine parameter based on the inspection results. The individual machines would still do the physical work, while the agent would operate at a higher level, deciding how the different pieces fit together.

 

Anthropic says MHS agents can sequence steps across instruments, monitor results and adjust parameters as conditions change. For long-running operations, the agent can also chain driver commands into code so hardware can execute the sequence without requiring the model to reason through every individual step.

 

This means an AI model does not need to sit in the control loop for every millisecond of a robot's movement. The low-level operation can remain deterministic, while the AI operates at a higher level, handling planning, monitoring and changes when something unexpected happens.

MHS in Physical AI

Physical AI systems need to be tested repeatedly. Developers need to try different robots, sensors, workflows and control strategies. If connecting each new piece of hardware requires weeks of engineering, experimentation becomes expensive.

 

Anthropic says its early MHS projects reduced the time required to integrate devices and allowed teams to iterate faster, while also helping with live machine operation and real-time fault detection.

 

More straightforward integrations mean more experiments. More experiments can mean more opportunities to develop reliable workflows. Those workflows can eventually be turned into repeatable automation.

 

Anthropic is already working with organisations across robotics, manufacturing and hardware. Amazon Web Services (AWS) plans to support MHS through Strands Robots, Hugging Face is adding support to LeRobot and Doosan Robotics is testing it with robotic arms, while Universal Robots has had early access and plans to add support to its robotics platform.

What MHS does not fix

None of this addresses the harder problems in Physical AI. A common driver does not make a robot better at handling objects with human-like precision. It does not close the gap between how objects behave in simulation and how they behave in the real world. Nor does it give an AI model a better understanding of friction, weight or the behaviour of unfamiliar objects.

 

Anthropic acknowledges that its models still have limitations in spatial and physical reasoning and says expert oversight remains necessary. In one example involving protein samples, Genentech researchers had to guide Claude to recognise that foaming was a physical problem rather than a software error.

 

There are also hardware limitations. MHS currently requires a programmable interface, meaning older or simpler equipment without one still needs additional work from manufacturers.

 

Safety is another major question. Giving an AI access to a machine is fundamentally different from letting it modify a spreadsheet or search a database. A mistake can damage equipment, ruin materials or create physical hazards.

 

MHS is also still a gated preview, not an open standard, and Anthropic says it wants to build safety evaluations with partners before changing that, following roughly the same path MCP took from its 2024 announcement to wider industry adoption.

 

If MHS does end up being widely adopted, the payoff is not a smarter AI. It is a developer being able to build an agent around a standard interface instead of starting the integration work over every time a piece of hardware changes.