Contents
Anthropic opened a research preview this week of something quietly bigger than another model bump: the Model Hardware Standard, or MHS, a shared specification that lets AI agents operate physical lab and manufacturing equipment through one consistent interface. It is built on top of the Model Context Protocol, the same open standard Anthropic shipped in 2024 for connecting agents to data sources and tools, except this time the “tool” on the other end can be a liquid handler, a robotic arm, or a laser on a quantum computing bench.
The problem MHS is aimed at is unglamorous but real. Getting instruments in a lab or on a factory floor to talk to each other, let alone to an AI agent, typically takes a specialist weeks or months of bespoke integration work, because every device ships with its own proprietary interface and nothing agrees on a common language. Anthropic built MHS with HHMI Janelia Research Campus specifically to collapse that integration time down to hours, and it is now sharing an early version with a first group of partners in biotech, robotics, and manufacturing, including Genentech and labs at the University of Washington, ahead of eventually open sourcing the standard entirely.
Three things worth understanding
The whole spec rests on two primitives: read and write. An MHS driver translates between the operating system and any device with a programmable interface using commands as simple as “read” (get the current temperature) or “write” (set the temperature). That deliberate simplicity is what lets one driver format cover a centrifuge, a microscope, and an industrial robot arm without a custom adapter for each.
Device knowledge comes from natural language, not just code. A lot of what makes hardware safe to operate, the weight of a robot arm, a torque limit, a calibration quirk, has historically lived in a paper manual or in someone’s head, not in machine-readable form. MHS drivers carry tags that let a person (or an agent interviewing that person) describe this in plain language, and the driver turns it into a reference file the agent can reason over before it ever sends a write command.
Agents get three ways to actually control the hardware. Once a device is described, an agent can drive it through MCP directly, through a command line interface, or by writing code files that chain driver commands into a deterministic script. Anthropic’s own testing found Claude using all three in sequence: probing a laser interactively through MCP like a scientist would, watching a camera feed to see what each adjustment did, then packaging what it learned into a script so the whole alignment routine could run as a single command without the agent re-reasoning at every step.
Try it: a simulated MHS console
About the demo
This is a static simulation of the read/write primitives MHS drivers expose, built from Anthropic’s own description of the protocol. It does not connect to any real device or to Anthropic’s infrastructure.Pick a device, an operation, and a parameter, then send a command.
Try writing dispense_volume_ul above its safety limit on liquid_handler_1 and the console rejects it the same way an MHS driver would refuse an unsafe command before it ever reaches real hardware.
Where this leaves developers
MHS is still a gated research preview, not a public API, so nobody outside the first partner cohort is wiring up a real centrifuge to Claude this week. But the design choice worth paying attention to is that Anthropic did not invent a new physical-world protocol from scratch. It extended MCP, the same standard already wired into countless SaaS tools and internal agents, and reused it as the transport for read and write primitives against hardware. If that pattern holds once MHS goes open source, the practical upside for anyone already building MCP servers is that the skills transfer directly: the same tool-definition and safety-boundary thinking that applies to a database MCP server applies to a device driver. My take is that the natural-language device tags are the more interesting bet long term, because they are Anthropic explicitly trying to digitize the tribal knowledge sitting in lab manuals and senior engineers’ heads before it is lost, not just wiring up an API. Worth applying for the preview at modelhardwarestandard.com if you run instruments an agent could plausibly supervise.