Most scientific labs are still duct-taped together with bespoke software, paper manuals, and specialist knowledge that lives in someone’s head. Getting a liquid handler to talk to a robotic arm, let alone hand control of both over to an AI agent, typically takes weeks of engineering work. Anthropic thinks it has a fix, and it’s starting to share it. The company announced a research preview of the Model Hardware Standard (MHS), a shared specification designed to let AI agents operate physical lab and manufacturing equipment through a single, standardized interface.
What MHS actually does
The core problem MHS solves is fragmentation. Every device in a lab or on a factory floor tends to have its own programming interface. There is no common layer for devices to share data with each other, and no standard way to expose them safely to an AI agent. MHS introduces a standardized driver that translates between hardware and software using a minimal set of primitives: commands like ‘read’ or ‘write’. That’s it. Any device with a programmable interface can understand those commands.
The driver also makes each device discoverable across a network in a standard format, so agents and devices can find each other without a custom translator sitting between them. And it includes a tagging system where users can write natural language descriptions of their hardware directly in the driver, covering things like the physical weight of a robot arm or the safety limits of a laser rig. That information gets compiled into a reference file the agent can read before it ever touches the device.
Once devices are connected, agents control them through three mechanisms: MCP (the Model Context Protocol), a command line interface, and code files via APIs. For long-running tasks, agents can chain driver commands together into scripts that run autonomously, without the agent reasoning through every individual step. During testing, Anthropic watched Claude align a laser by adjusting it, observing results through a camera, iterating, and then packaging what it learned into a deterministic script it could reuse.
Who’s already using it
Anthropic shared early versions of MHS with partners across biotech, robotics, and quantum computing before this preview. The results are worth paying attention to:
- Genentech used MHS to automate a BCA protein assay, coordinating a liquid handler, robotic arm, and plate reader as a proof-of-concept.
- A PhD student at the University of Washington Baker and Pinglay labs built a remote instrument dashboard, an agent-supervised qPCR workflow, and a collision-free integration between a robotic arm and liquid handler.
- Carnegie Mellon researchers ran serial dilution dose-response experiments roughly three times faster than before, with an agent orchestrating four devices spread across three computers with incompatible interfaces.
- HHMI Janelia, where MHS development started as a collaboration with Anthropic, used it to unify a microscopy rig that previously required seven separate vendor programs.
- QuEra, a neutral-atom quantum computing company, gave an AI agent control over parts of their laser stabilization system.
Why this matters beyond the lab
MHS is model-agnostic and works with any agent framework that supports standard protocols like MCP. That’s a deliberate choice. Anthropic is not trying to lock this to Claude. The goal appears to be establishing the standard itself, which is a smarter long-term play than building a proprietary integration layer that competitors would ignore.
The comparison that comes to mind is USB. Before it, connecting peripherals was a mess of incompatible ports and drivers. MHS is attempting something similar for scientific and industrial hardware, at a moment when the appetite for agentic AI in physical environments is growing fast. Competitors like Microsoft and Google are pushing hard into agentic workflows, but neither has made a move this specific toward physical device control at the lab and factory level.
Anthropic is sharing MHS with a first group of scientific research labs and advanced manufacturers ahead of open-sourcing it, with the stated aim of building safety evaluations and best practices for AI operating physical equipment before releasing it broadly. That sequencing is the right call. An AI agent misreading a temperature parameter in a drug discovery workflow is a different category of risk than a chatbot giving a bad answer.
If MHS gains adoption, the practical effect is that research teams spend less time on integration work and more time running experiments. For labs where instrument setup currently takes weeks, that is a real operational shift. Whether the standard spreads beyond early adopters will depend on how well Anthropic manages the open-source release and whether hardware manufacturers have any incentive to support it natively. That part remains to be seen.
