Is the Model Hardware Standard open source?

Not yet. It’s currently a limited research preview available to a small group of labs and manufacturers by waitlist ; Anthropic plans an open-source release, with developer documentation, once its safety evaluations are further along.

How is MHS different from the Model Context Protocol (MCP)?

MCP standardizes how AI models connect to software tools and data. MHS extends the same idea to physical hardware, adding device discovery and safety metadata, and can itself be accessed through MCP as one of its control paths.

Which labs are already using the Model Hardware Standard?

Early testers include Genentech, Carnegie Mellon University, University of Washington labs, HHMI Janelia Research Campus, QuEra Computing, and Tetsuwan Scientific, alongside hardware makers like Danaher, Tecan, QIAGEN, and Universal Robots adding support.

Can I sign up to try it myself?

Only if you’re part of a research institution or manufacturer — access is by application through the official waitlist , not a public signup.

Is it safe to let AI control physical lab robots?

Anthropic enforces safety limits at the device-driver level and says it’s building fuller physical-safety evaluations with its launch partners, but it’s explicit that Claude’s spatial reasoning still has limits and that expert human oversight remains necessary.

What Is the Model Hardware Standard — AI That Runs Lab Gear

The Model Hardware Standard (MHS) is a technical specification that lets AI agents safely operate physical lab and manufacturing equipment — liquid handlers, robotic arms, microscopes, centrifuges — using one shared set of commands instead of custom-built code for every machine. Anthropic introduced it as a research preview for scientific labs and advanced manufacturers, extending the same idea behind the Model Context Protocol — its standard for connecting AI to software tools and data — into the physical world.

The Problem It’s Trying to Solve

Scientific instruments and factory robots are built by dozens of vendors, each with its own control software, communication protocol, and quirks. Wiring a new device into an automated workflow has traditionally meant writing a bespoke “translator” program — a task that Laboratory automation specialists say can take weeks or months per instrument, and that has to be redone whenever a lab swaps in a different robotic arm, reader, or sensor. That integration bottleneck is a big reason labs still automate only a fraction of their repetitive work, even when the hardware to do it already exists on the bench.

How It Works

MHS standardizes hardware access around a small set of ideas:

Read and write primitives. Every device exposes simple actions like “get temperature” or “set temperature” instead of a proprietary command set, so an AI agent doesn’t need device-specific code to ask a plate reader for a measurement or tell a robotic arm to move.
Device discovery. Equipment announces itself on the network in a standard format, so an agent can find and address it without a hand-written driver.
Safety metadata. Each device carries natural-language tags describing its weight, movement limits, and hazards — information that used to live only in a PDF manual or in a technician’s head — so an agent can check what’s safe before issuing a command.
Three ways in. Agents can control devices through MCP, a plain command-line interface, or generated code files for multi-step workflows, and the standard is model-agnostic — any AI system can use it, not just Claude.

Together, this lets a researcher describe a procedure in plain language — “run this assay across the liquid handler, then move the plate to the reader” — instead of stitching together separate control scripts for each instrument.

Who’s Already Using It

Anthropic opened the research preview to a small group of labs and manufacturers. Genentech used it to automate a protein assay across a liquid handler, robotic arm, and plate reader in one workflow. Carnegie Mellon ran dose-response experiments roughly three times faster. University of Washington labs used it to monitor lab equipment and hand off plates between instruments automatically. HHMI Janelia Research Campus orchestrated microscopy equipment that normally requires seven separate vendor programs, and QuEra Computing applied it to laser control on quantum hardware, reporting a 99.3% recovery rate after interruptions. Hardware makers including AWS, Danaher, Doosan Robotics, Tecan, QIAGEN, and Universal Robots are building in support.

Why It Matters

MHS is part of a broader shift of AI agents from software tasks into direct control of physical equipment — the same trajectory driving interest in robotics and lab and factory automation more broadly. If a common hardware standard catches on the way MCP did for software tools, it could lower the cost of automating routine scientific and manufacturing work considerably, letting smaller labs run experiments that today require dedicated automation engineers.

That said, Anthropic is explicit that the standard is still a research preview, not a finished product. Claude and other AI models learn about the physical world through text and images, so their spatial and physical reasoning has real limits — a wrong command sent to a real robotic arm can break equipment or cause injury in ways a wrong command to software cannot. Anthropic says device-level safety limits are enforced through MHS’s drivers, and that it’s building out safety evaluations and a fuller physical-safety framework with its launch partners before any open-source release, alongside guidance for safe deployment. Expert human oversight is still required.

How to Try It

Access is currently limited to research institutions and manufacturers through a waitlist, rather than a public download. Anthropic says it plans to open-source the standard, along with developer documentation and findings from the preview, once its safety evaluations are further along — there’s no public release date yet.

In the news

Anthropic announced the Model Hardware Standard as part of a research partnership with HHMI Janelia and a first cohort of labs and equipment makers testing it under real lab conditions.