Anthropic Opens MHS, a Standard for AI Agents to Operate Lab and Factory Equipment

When a rumor circulated last month that Anthropic was acquiring robot AI company Physical Intelligence, the path the company actually took was not an acquisition. Rather than standing up its own hardware lab the way OpenAI has, it chose to build the software layer that sits on top of robot hardware. What that strategy looks like in practice has now arrived.

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices. It lets agents run microscopes, liquid handlers and robotic arms in parallel, handling work that ranges from routine drug discovery experiments to laser calibration on a quantum computer. The standard began as a collaboration between Anthropic and HHMI Janelia Research Campus.

Months of Integration Work, Cut to Hours

Setting up and integrating hardware in a lab or manufacturing facility typically takes weeks, often months. Devices each carry their own programming interface, so specialists build bespoke integrations every time. Even once connected, there has been no common way for those devices to share data with an AI agent or to let the agent operate them safely.

MHS addresses this with a standardized driver built on simple primitives — “read” (get temperature), “write” (set temperature) — that any device can act on. Each device becomes discoverable in a standard format, so devices and agents can find each other across networks without a bespoke translator in between.

The more interesting piece is that information code cannot express gets written in natural language. The weight of a robot arm, for instance, matters for handling it safely but has historically lived in paper manuals, on someone’s laptop, or as tacit knowledge. Users can write it themselves or have an agent interview them about their setup. From those tags the driver automatically produces a reference file describing what the device can measure, what can be adjusted, and what safety limits will be enforced.

Control runs through three mechanisms — MCP, the command line and code files — which together allow orchestration across multiple devices from a single line of code. For tasks too long or too fast for step-by-step reasoning, the agent chains driver commands into code files so the devices execute on their own.

Anthropic reports that in testing, Claude approached hardware the way a scientist would. It adjusted a laser, watched through a camera to see how the adjustment moved the beam, and repeated the process. It then packaged what it learned into a deterministic script that aligned the laser as a single command, without reasoning at each step.

Not the Brain — the Layer Beneath It

To see what MHS is actually going after, it helps to separate the physical AI stack into layers.

At the top sit orchestration agents like Claude. Below them are robot foundation models: NVIDIA’s GR00T N1.7, a vision-language-action model that takes images and language and produces actions across everything from tabletop arms to humanoids; Physical Intelligence’s π0; Skild AI’s Skild Brain; Generalist‘s GEN-1.5. Beneath that is the device abstraction layer, and beneath that the hardware itself.

MHS targets the device abstraction layer. It is not trying to build a brain. It is building the path a brain has to travel to move real equipment — which makes GR00T and π0 potential consumers of MHS rather than competitors to it. NVIDIA’s Isaac is a full-stack platform bundling simulation (Isaac Sim and Lab), middleware (Isaac ROS) and on-robot inference hardware (Jetson Thor), but it is designed around NVIDIA’s own chips and robots, which makes it a different kind of thing.

The Real Competition Is Lab Standards

The standards MHS actually collides with are elsewhere, and there are several of them.

SiLA 2, the lab automation standard, is built on web technologies and is generally considered approachable for IT teams. The OPC UA camp, rooted in process industry automation, has put forward LADS as a laboratory device extension and now competes with SiLA 2. Robotics has ROS 2 as its de facto middleware. Layered on top of all that are legacy industrial protocols like Modbus and PLC, plus older serial-connected instruments, often mixed together inside a single lab.

Integration is hard not because standards are missing but because there are too many of them. That leaves MHS exposed to a familiar trap: competing standards proliferate, someone builds a unifying standard, and now there is one more. MCP became a standard quickly in part because the position was essentially vacant. The device abstraction layer is not vacant.

There is still a genuine gap MHS is aimed at. SiLA 2 and OPC UA were designed for machines to talk to machines, not for an LLM agent to make sense of equipment it has never encountered. Writing down in natural language how heavy a robot arm is and which operations are dangerous, so an agent can read it, is not something the existing standards do.

A Company That Has Already Pulled This Off Once

Anthropic has run this play before. MCP, released in November 2024, standardized how AI reaches external data and tools. OpenAI adopted it in March 2025, with Google DeepMind and Microsoft following mid-year. That December, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, with Google, Microsoft and AWS signing on as supporting members. It now sees roughly 97 million monthly SDK downloads across some 10,000 active servers.

MHS carries the same design intent. It is model-agnostic rather than Claude-only, any agent harness can reach it through standard protocols such as MCP, and Anthropic has signaled it will be open-sourced. Making it easy for competitors to adopt looks like a concession, but that is precisely how MCP became the standard: don’t sell the hardware, build the socket everyone has to plug into.

Standards contests, though, are settled by distribution rather than by specification quality — by what device manufacturers ship in their products by default. Read the partner list through that lens and the shape becomes clear. Danaher, QIAGEN and Tecan are giants of laboratory instrumentation. Automata is adding MHS to its lab automation platform, and MBF Bioscience is building a driver for ScanImage, the software running laser-scanning microscopes in hundreds of neuroscience labs. AWS will support MHS through Strands Robots, Hugging Face through its LeRobot robotics library, and Raspberry Pi across several of its products. Large research facilities and hobbyist hardware, seeded at the same time.

Doosan Robotics Is In

One Korean company appears on the partner list. Doosan Robotics is testing MHS with its robotic arms, including automated quality assurance and coordinating tasks across multiple robots. Universal Robots, the other major force in collaborative robotics, has had early access and plans to add support to its platform — putting both leaders of the cobot market on the list.

For Korean robotics companies, a settled standard here lowers the barrier. Connecting a robot to AI has meant custom integration per vendor; a single driver would let any agent attach. The counterweight is that Anthropic leads the standard, which leaves a longer-term dependency question. Whoever writes the specification also decides what counts as a safety limit and which devices get supported first.

The Limits Are Explicit

Anthropic is clear that the standard is unfinished. Claude learns about the physical world through text and images, so its spatial and physical reasoning has limits that still require expert oversight. Working with protein samples, Genentech researchers had to guide Claude to recognize that errors caused by foaming were physical failures rather than software bugs, fixable only through physical correction.

Devices without a programming interface are not supported yet; Anthropic says it is working with those manufacturers to build drivers. During the preview it plans to add safety evaluations and develop a physical safety roadmap for AI operating in the physical world.

An AI writing text on a screen and an AI firing a laser in a lab carry different kinds of risk. That the party defining the standard also proposes to define the safety bar follows naturally — and raises the question of who sets that bar.