Anthropic previews MHS, a standard for AI agents to control lab and industrial hardware

The Model Hardware Standard extends the same idea as Anthropic's Model Context Protocol into the physical world, but instead of standardising how models reach data sources it standardises how they reach devices. Drivers expose equipment through a small set of primitives, read and write among them, and attach structured tags that describe what a device can do, including physical characteristics such as a robot arm's weight and range. Setup is conversational: a user tells the model what is connected, and the driver produces a reference file the agent can consult when it encounters hardware it has never seen. The underlying assumption is that language models already know enough about microscopes and robot arms from text and images to plan around them, even though they lack physical intuition about force, reach or collision risk.

Once connected, an MHS-equipped agent can issue commands through three routes: the Model Context Protocol, a command-line interface, or generated API code. Anthropic's demos include adjusting a laser, inspecting the result through a separate camera and iterating to calibrate the system, as well as focusing a microscope, analysing the image, picking a region of interest and moving the stage to it. In one demo, Claude reportedly figured out how to get a robotic arm to pick up an aluminium can without task-specific training. The API-script path is the one that matters most for real workflows, because it lets the model sequence actions across several devices and edit the sequence as conditions change rather than reasoning through every step from scratch.

Anthropic cites two concrete results from the preview: Genentech running a drug-discovery experiment with real-time error handling, and quantum-computing company QuEra improving laser-stabilisation performance from 58% to 99.3%. The announcement does not include independent benchmarks, the test conditions or how much human oversight each setup required, so the numbers are vendor-supplied until reproduced. The QuEra figure in particular is striking and should be treated as a single-shot claim until the methodology and supervision regime are published.

The initial test site is the Howard Hughes Medical Institute's Janelia Research Campus in Maryland, and other labs or manufacturers with suitable equipment can apply to join the preview. Anthropic's partner lists are inconsistent across the materials supplied: one names AWS Strands Robots, Automata LINQ, Danaher, Doosan Robotics, MBF Bioscience, Qiagen, Tecan and Universal Robots, while another names AWS, Hugging Face's LeRobot, Raspberry Pi, Automata and Universal Robots. There is no consumer product, no public open-source release and no pricing attached to today's announcement, so the reachable audience is narrow by design.

The headline benefit is integration time, which Anthropic says drops from weeks or months to hours or minutes when devices speak MHS. The headline risk is the same: a thinner integration path is also a thinner barrier between a language model and a piece of moving equipment, and the standard's safety tags are now the load-bearing wall. Adoption will hinge on whether the reference-file model produces reliable behaviour on devices the agent has never seen, and on whether the eventual open-source release ships with evaluations that are independent of Anthropic's own demos.