AI • 4 min read
Anthropic is previewing MHS, a hardware interface that lets AI agents operate lab equipment, robots and factory systems—but safety work comes first.
Image: The Register Anthropic is testing a new interface that could let AI agents operate laboratory instruments, industrial equipment and robots without a custom integration for every device. The Model Hardware Standard (MHS) is now in a research preview, with Anthropic targeting a future open-source, agent-agnostic release rather than shipping a finished developer standard today.
MHS is conceptually similar to Anthropic’s Model Context Protocol, but it connects models to physical systems instead of data sources. The company describes the project in its official research-preview announcement as a way to let models such as Claude “safely operate physical devices.” That safety claim is still a work in progress: Anthropic says the preview is intended to generate more evaluations and protections for systems acting in the physical world.
The immediate US relevance is limited. The preview is aimed at scientific research labs and advanced manufacturers with suitable equipment, and there is no consumer product, public open-source implementation or pricing attached to the announcement.
Recommended reading
Bill Gates says AI has crossed five danger thresholds
Ava Chen • • 7 min read
A translation layer for incompatible machines #
Most lab and manufacturing equipment exposes some form of programmable control, but the interfaces, terminology and constraints vary widely. Anthropic says integrating a facility’s hardware can take weeks or months because specialists must build bespoke connections between machines that were never designed to work together. Its stated goal is to reduce that work to hours or minutes.
MHS drivers expose devices through a small set of primitives, including “read” and “write.” The driver makes connected equipment discoverable in a standard format and adds tags describing what each device can do. Those tags can encode physical characteristics such as a robot arm’s weight and range, along with adjustable parameters, measurement options and enforced safety limits.
During setup, users can provide device information through a conversation with the model. The driver then produces a reference file that gives an agent structured information about hardware it may never have encountered during training. That matters because a language model can describe a microscope or robotic arm from text and images without having reliable physical intuition about force, reach, collision risk or the consequences of an incorrect setting.
An MHS-connected agent can control equipment through three paths: the Model Context Protocol, a command-line interface or API code. It can issue commands, monitor test results and change instrument settings. Anthropic also says models can write API scripts to sequence operations across several devices, then modify those scripts as conditions change rather than reasoning through every individual action from scratch.
The examples are operational. A model could adjust a laser, inspect the output through a separate camera and repeat the process to calibrate the system. It could focus a microscope, analyze an image, identify an area requiring closer observation and move the instrument to that section. In another demonstration, Claude reasoned through how to make a robotic arm pick up an aluminum can without being specifically trained on those steps.
Early results and the safety boundary #
Anthropic says biotechnology company Genentech used MHS to run a drug-discovery experiment with real-time error handling. Quantum-computing company QuEra used it for laser stabilization, where Anthropic says the system improved the result from 58 percent to 99.3 percent. The announcement does not provide independent benchmarks, the test conditions behind those figures or details about how much human supervision each system required.
The initial test site is the Howard Hughes Medical Institute’s Janelia Research Campus in Maryland. Anthropic says other organizations with suitable laboratory or industrial equipment can apply to participate in the research preview. The partner lists supplied with the announcement vary: one includes AWS’s Strands Robots library, Automata’s LINQ platform, Danaher, Doosan Robotics, MBF Bioscience, Qiagen, Tecan and Universal Robots; another identifies AWS, Hugging Face’s LeRobot, Raspberry Pi, Automata and Universal Robots as part of the first group of scientific and manufacturing partners.
“It typically takes a lab or manufacturing facility weeks, if not months, to set up and integrate their hardware. Most devices don’t communicate with each other, instead requiring specialists to build bespoke integrations. MHS reduces this integration work to hours or minutes.”
That reduction in integration time is also the main technical risk. A standard that makes it easier for an agent to issue commands across instruments could make experiments faster, but it could also turn a mistaken assumption about a device’s limits into a physical failure. Anthropic acknowledges that large language models learned about the physical world primarily from text and images and says it needs more testing before releasing the system as open source.
“If you can test hypotheses faster, you could create general technologies faster. This is how a century of progress can condense into a decade.”
For now, MHS is integration plumbing under controlled evaluation, not an autonomous lab operating system.
[Ava Chen](/authors/ava-chen/)
AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.