Anthropic opens Model Hardware Standard research preview for physical agents

Anthropic opened a research preview of the Model Hardware Standard on August 27, 2026. MHS is a shared specification for AI agents to operate programmable lab and manufacturing devices—not a generally available robot, and not an Anthropic-only stack.

The company is letting a first group of scientific research labs and advanced manufacturers try an early version. Agents are meant to run multiple instruments at once, including microscopes, liquid handlers, and robotic arms, on work that spans routine drug-discovery experiments and laser calibration on a quantum computer. Reuters independently reported the same preview and the same device classes.

Development started as a collaboration between Anthropic and HHMI Janelia Research Campus. Anthropic says MHS is model-agnostic: any agent harness can reach it through standard protocols such as the Model Context Protocol. The plan is to open-source the standard after partners help build safety evaluations and best practices. Until then, access is by application.

That is the product shape practitioners should plan around. A coding-agent harness that already speaks MCP is being asked to drive real gear through a common driver. Integration time, eval design, and physical safety all move with that interface.

Hours instead of weeks of glue code

Anthropic's diagnosis is familiar on any lab floor: most devices do not communicate with each other, so specialists spend weeks or months on bespoke integrations. MHS inserts a standardized driver that translates between a computer's operating system and a hardware device.

The driver uses simple primitives—read (for example, get temperature) and write (for example, set temperature)—that any programmable device can act on. It also makes each device discoverable in a standard format so instruments and agents can find each other across a network without a custom translator.

Physical facts that never lived in an API are part of the contract. The driver includes tags so a user can write, in natural language, characteristics such as the weight of a robot arm. Those tags produce a reference file: what the device can measure, what can be adjusted, and which safety limits will be enforced. Ars Technica describes the same tagging layer as a way to encode real-world constraints for models trained mostly in software.

Control then happens three ways that Anthropic says work together: MCP, a command-line interface, and code files (APIs). For long-running tasks, or work faster than online reasoning allows, an agent can chain driver commands into scripts the devices run themselves.

In testing, Anthropic says Claude treated hardware the way a scientist might: adjusting a laser, watching the beam through a camera, repeating the loop, then packaging the result as a deterministic script. The company is also clear about failure modes. MHS does not yet work on hardware without a programming interface. Spatial and physical reasoning still need expert oversight. Genentech researchers, automating a BCA protein assay across a liquid handler, robotic arm, and plate reader, had to teach the model that foaming errors were physical failures, not software bugs.

Who is in the preview

The public Model Hardware Standard site calls this a limited research preview. Stakeholders in science and industry can apply to test the standard, build safety evaluations, and develop practices for agents on physical equipment before open source.

Anthropic lists early scientific work from Genentech; the University of Washington Baker and Pinglay labs, including remote monitoring, agent-supervised qPCR, and collision-free plate handoffs; Carnegie Mellon, where serial dilution dose-response runs ran about three times faster across three computers with incompatible interfaces; HHMI Janelia, unifying a microscopy rig that previously needed seven vendor programs; QuEra Computing, where an agent recovered laser lock 99.3 percent of the time without human intervention; and Tetsuwan Scientific, running qPCR to profile pollution in San Pedro Creek.

Hardware and platform partners named on Anthropic's announcement and in Ars Technica's report include Amazon Web Services, which will offer a private pre-release of Strands Robots; Hugging Face, adding MHS support in LeRobot; Raspberry Pi, after tests with a Camera MHS Driver; Universal Robots, with early access and plans to add support; plus Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, and Tecan.

Anthropic says it will use the preview to strengthen a physical safety roadmap and that preview findings will accompany guidance when MHS is open-sourced.

The SO-ARM101 clip is the same brief

Hugging Face's LeRobot account posted the same day that it took part in the MHS research preview. The post shows Claude Code running a real SO-ARM101. "Nothing in this was trained," LeRobot wrote—"no policy, no teleoperation, no demonstrations." The agent measured the workspace itself and wrote the motion.

That demo is a partner preview, not general-availability hardware support and not a second launch. It is useful because it shows the intended loop: a coding agent, a shared spec, a programmable arm, and no trained policy. If you evaluate physical agents, the unit under test is that loop—discovery, reference tags, motion code, and the human who still has to catch a physics mistake the model will not see.

If you run programmable lab or factory equipment, apply for the research preview. If you build agent harnesses, read Anthropic's driver write-up next to LeRobot's SO-ARM101 post and treat MCP plus the shared driver as the integration surface—not a robot SKU, and not general availability.

Hours instead of weeks of glue code

Who is in the preview

The SO-ARM101 clip is the same brief

Sources