Anthropic Moves Into Physical AI With Model Hardware Standard

Anthropic is making a clear move from software agents into Physical AI. On August 27, the company introduced Model Hardware Standard, a specification intended to standardize how AI agents discover devices, understand their constraints, issue commands and receive machine state. The idea resembles an integration layer for software, except that failures now have physical consequences.

The first preview is being shared with scientific labs and advanced manufacturers. Anthropic lists microscopes, liquid handlers, robotic arms and equipment used in biology and quantum systems. MHS is described as model-agnostic and can be accessed through standard mechanisms including MCP.

What MHS actually does

The core concept is a standardized driver that describes a device and exposes common primitives such as read and write, together with state and operating limits. The driver can also carry information that cannot be inferred safely from an API alone, such as a robot arm’s weight, movement constraints or other safety-relevant characteristics.

Once devices are connected, an agent can orchestrate them as one workflow: run a liquid handler, ask an arm to move a plate, read a measurement and change the parameters of the next experiment. Anthropic is therefore positioning MHS as an orchestration layer for fleets of heterogeneous equipment, not merely as a way to make Claude move one robot arm.

The early tests are meaningful, but still experimental

Anthropic describes a Genentech proof of concept in which Claude used MHS to coordinate a liquid handler, robotic arm and plate reader. The system optimized liquid-transfer parameters and recovered from some errors. The company also documents work in academic laboratories, microscopy and other automation settings. Closed-loop measurement, decision and action is stronger evidence than a stage demonstration, but it is still early research evidence.

The same material also exposes current limitations. In a protein-handling experiment, the model initially failed to understand the physical cause of bubble-related errors and needed expert guidance. That distinction matters: a communication standard does not solve physical reasoning. MHS may make hardware easier for an agent to access, but it does not guarantee that the model understands the physics behind a failure.

Why this matters for robotics

Physical AI remains fragmented. Robot makers expose different SDKs, control stacks and data formats, while integrators repeatedly build bespoke glue. If MHS or a competing standard wins broad support, part of the industry’s value could shift away from a single robot body toward the layer that allows one agent to work across many machines.

This is also an interesting counterpoint to the humanoid race. A humanoid tries to make one body compatible with environments built for people. MHS attacks the problem from the other direction: it does not require one universal body, but instead tries to give an AI agent a common language for many specialized devices.

What remains a manufacturer claim

• Anthropic says integrations that currently take weeks or months can in some cases be reduced to hours or minutes. This is not an independent industry-wide benchmark.
• The company describes autonomous, round-the-clock workflows and recovery from some hardware errors. The degree of autonomy depends on the specific system and task.
• Anthropic intends to open-source MHS after the preview. The complete public release is not yet available at publication time.

Europe: a hardware standard meets machinery regulation

For European cobot, automation and laboratory-equipment companies, MHS could be strategically relevant but also regulatory-sensitive. The EU Machinery Regulation 2023/1230 applies from January 20, 2027 and explicitly addresses certain AI-based safety functions and machinery with evolving behavior. A layer that describes motion limits and lets an agent operate physical machinery therefore needs safety to be architectural, not an afterthought.

RoboMorrow view

MHS is not yet a universal robot operating system, but it could become an important piece of the Physical AI stack. The key story is not that Claude moved a robotic arm. It is that Anthropic is trying to standardize the interface between AI models and the physical world. If hardware vendors begin implementing such a layer natively, the integration barrier for machine-operating agents could fall sharply.

MHS is not a robot brain — it sits at a different layer

It is easy to mistake MHS for another VLA or robot foundation model. That would be wrong. A motion or action model decides how a robot should move a hand, leg or gripper. MHS is trying to standardize the integration layer below that: how an agent discovers a device, learns its states and operations, reads its constraints and sends commands. In one system the device may be a robotic arm; in another it may be a microscope, pump, furnace or camera.

That distinction is strategically important. The market can support many competing “brains” while all of them benefit from hardware speaking a common language. Anthropic describes three control mechanisms: MCP, a command-line interface and code/API files. An agent can use reasoning to design a task, then package repeatable steps into deterministic code instead of asking an LLM to reason through every millisecond of machine operation. That separation matters for latency, repeatability and safety.

The early numbers are impressive — and partner-reported

The most concrete case study comes from QuEra Computing, which used MHS while automating recovery of laser locks inside neutral-atom quantum systems. According to the partner material published by Anthropic, a bespoke recovery script built over several months succeeded roughly 58% of the time and took about 150 seconds per attempt. The MHS-based system later recovered 695 of 700 blind-test disturbances, or 99.3%. Simple recoveries reportedly took 0.9–5.4 seconds and harder disturbances roughly 10–14 seconds. These are useful data points, but they are not an independent RoboMorrow benchmark.

The Genentech example is equally valuable because it shows failure. Claude could optimize liquid-transfer parameters and recover from some runtime errors, yet it initially treated foaming and bubbles as if they were software problems. Experts had to explain the physical cause and the appropriate correction. Anthropic publishes this limitation directly. A standardized hardware interface does not magically give a language model physical, chemical or process intuition.

This is already broader than an Anthropic-only experiment

The preview includes companies across several layers of the stack. Universal Robots plans support on its robotics platform, Doosan Robotics is testing MHS with robotic arms, Hugging Face is adding MHS support to LeRobot, and Raspberry Pi is enabling integrations across a number of products. AWS is supporting preview participants through Strands Robots, while Automata, Danaher, QIAGEN and Tecan are exploring laboratory automation use cases. That is not mass adoption, but it is broad enough to make MHS an ecosystem attempt rather than a single Claude demo.

Hardware vendors now have to decide whether maintaining an MHS driver alongside their own SDK creates enough demand to justify the cost. Integrators need evidence that the common layer reduces engineering effort rather than simply adding another abstraction. Industrial customers will care about a different checklist: authentication, permissions, audit logs, deterministic safety limits, emergency-stop behavior and certification.

RoboMorrow Reality Check

The biggest opportunity for MHS is not flashy robot control. It is lowering the integration cost of Physical AI. If a device can describe its capabilities, limits and state to an agent in a standard way, building multi-machine workflows becomes much easier. But the current preview does not yet establish long-duration reliability, cybersecurity, hard fail-safe behavior, responsibility for unsafe commands or behavior during network loss.

MHS should therefore be treated as a serious candidate standard, not a finished “control layer for the physical world.” The next test is adoption in shipping hardware and independent reproduction: can outside teams achieve the promised integration savings while maintaining industrial-grade safety and reliability?

Sources

Primary source: Anthropic — Model Hardware Standard research preview. Independent confirmation: Reuters. Official standard site: Model Hardware Standard.