Anthropic's Model Hardware Standard gives AI agents a common interface to control physical lab and factory equipment

Plug a microscope, a robotic arm, and a liquid handler from three different vendors into the same lab bench and you will almost certainly need three different software stacks to operate them. Getting an AI agent to orchestrate all three in a single workflow has, until now, meant months of custom integration work. Anthropic’s answer to that problem is the Model Hardware Standard (MHS), which entered research preview on 27 August 2026.

What MHS actually is

MHS is an open, model-agnostic protocol that gives AI agents a common interface to discover physical hardware, read from its sensors, send commands to its actuators, and respect its safety limits, all through a standardised driver format. It is built on top of MCP (Model Context Protocol), the specification Anthropic open-sourced in 2024 to standardise how AI models connect to software tools and data sources.

Think of MCP as the layer that connects AI to digital systems. MHS extends that same logic one layer further into the physical world. Each device exposes a set of common commands (“read”, “write”) and an MHS file that describes its operational parameters: speed limits, weight constraints, emergency-stop conditions, and so on. An AI agent can read that file, understand what the device is capable of, and start working with it without any bespoke integration code. Anthropic describes it as the “USB moment” for physical AI, and that analogy is reasonably apt. USB did not make all peripherals identical; it gave them a common plug so your computer could negotiate with any of them. MHS does the same thing for lab and factory equipment.

Where it came from

The project began as a collaboration between Alek Kemeny on Anthropic’s Beneficial Deployments team and Arco Bast, a postdoctoral scientist at HHMI Janelia Research Campus. Bast was running complex brain-imaging experiments on a rig combining lasers, motorised focusers, and specialised cameras from different vendors, with no common interface between them. He built a shared memory dictionary so the instruments could talk to each other at memory speed. Kemeny worked with him to layer AI models on top of that interface, and MHS grew from there.

Early results from the research preview

Anthropic has shared MHS with a first group of partners in science, robotics, and manufacturing. The numbers they are reporting are striking.

QuEra, which works on quantum computing hardware, tested MHS on laser relock tasks, the fiddly process of realigning a laser that has drifted out of calibration. The prior benchmark success rate was 58%. With an MHS-connected Claude agent handling the task, that rose to 99.3%. In safety testing, MHS blocked all six induced fault conditions before any hardware moved.

At Genentech, researchers automated a BCA protein assay across a liquid handler, robotic arm, and plate reader. A Claude Opus 4.8 agent ran serial dilution experiments three times faster than the previous baseline. Researchers did have to guide Claude to recognise that errors caused by foaming in samples were physical failures rather than software bugs, which points to a real limitation: LLMs learn about the physical world through text and images, and their spatial and physical intuition still has gaps that require expert oversight.

Anthropic also claims MHS reduces the time to integrate complex machinery with AI from months down to hours or minutes, which, if it holds across a wider range of hardware, would remove a significant bottleneck for any organisation wanting to run agentic scientific workflows.

What this means for researchers and engineers

If you work in a lab or a manufacturing environment, the practical implication is that you may eventually be able to connect a new piece of equipment to an AI agent the way you currently connect a printer to a laptop: plug it in, let the standard handle the negotiation, and start working.

That matters because, as Jonah Cool of Anthropic’s life sciences arm put it, “In many cases, the science doesn’t happen because you can’t use the equipment.” Instrument integration is a genuine bottleneck in scientific research, and it is often solved by graduate students writing one-off scripts rather than by anything reusable. MHS is an attempt to make that layer of work disappear.

For engineers in manufacturing, the same logic applies. Amazon Web Services has said it plans to support MHS through its Strands Robots library. Doosan Robotics, Universal Robots, Tecan, and QIAGEN are exploring integrations. If the standard achieves broad adoption, swapping one vendor’s robot arm for another’s becomes considerably less painful.

Important caveats

MHS is in research preview, not general availability. It is currently limited to a select group of partners, and Anthropic plans to open-source the standard only after early safety evaluations are complete.

There are technical limitations worth knowing about. MHS does not yet work with hardware that lacks a programmable interface, and Anthropic is working with device manufacturers to build in drivers for such equipment. The LLM reasoning layer also still requires human oversight for physically ambiguous situations, as the Genentech foaming example illustrates.

There is also a regulatory dimension. The EU’s AI Act machinery provisions take effect from January 2027, and an MHS file that constrains a robot arm’s speed and range of motion is, in regulatory terms, performing a safety function. For high-risk equipment categories, that means self-declaration of conformity may no longer be sufficient. Organisations planning to use MHS in production machinery should factor that into their compliance planning well before that date.

The broader picture

MHS represents Anthropic’s first serious move into physical AI, a domain where OpenAI and Amazon have already committed substantial investment. Anthropic has also been building out a silicon team to design custom chips, and hired Caitlin Kalinowski, a hardware executive previously at OpenAI, Meta, and Apple, signalling that the company is taking the physical layer seriously rather than treating it as an afterthought.

For anyone building or operating AI-connected scientific or industrial equipment, MHS is worth watching closely. The research preview is limited for now, but an open-sourced version of this standard, backed by a growing list of device manufacturers, would meaningfully change how quickly agentic workflows can reach physical hardware.