Anthropic unveils MHS, a standard for AI-controlled lab hardware

Claude steps outside the screen

On August 27, 2026, Anthropic opened a research preview of the Model Hardware Standard (MHS), a framework meant to let AI agents safely operate physical equipment such as microscopes, liquid handlers, and robotic arms. It isn't built for any single Claude model — it's a shared standard meant to connect any device with a programming interface, regardless of manufacturer or AI model. Anthropic says the goal is to cut equipment integration work that has historically taken weeks or months down to hours or minutes.

The Model Hardware Standard is a set of rules that defines how AI agents should — and should not — interact with all kinds of hardware. It reflects a growing belief that if AI can safely step into the physical world, it has the potential to transform fields like scientific research and manufacturing.

Anthropic says that before releasing the standard publicly, it plans to work with trusted partners to figure out how to maximize safety. There are potential misuse risks — developing biological weapons, for instance — but the company argues that safeguards built into the AI models themselves should be enough to stop bad actors from turning the new standard toward harmful ends.

"The starting point was wanting to accelerate scientific progress," said quantum physicist Alec Kemeny, who co-led the development. "How do you close the loop from accelerating literature review and data analysis to actually bringing that power into the experimental world?"

Claude and other chatbots are already powerful tools for scanning the vast body of information locked in scientific papers and experimental results to surface new insights and ideas.

AI agents are widely seen as the next step beyond chatbots. They're typically designed to take actions on a regular computer, like responding to email. But they could plausibly use other hardware too, and Anthropic wants rules in place before that happens.

The first official video shows MHS linking different pieces of lab and factory equipment under a single agent. Rather than stopping at calling tools on a screen, the AI moves on to reading sensors and sending commands to robotic arms and lab equipment directly.

MHS is the USB-C of hardware

Lab equipment looks similar on the surface, but each device speaks its own language. Microscopes, plate readers, and liquid handlers all have different programming interfaces, so linking multiple devices for a single experiment used to require an expert to build a custom translator. Even after devices were connected, there was no common way for an AI agent to read data and issue commands safely.

MHS inserts a standard driver layer in between. The driver provides basic commands that any device can understand in common — read to pull a temperature, write to set one. A device's weight, the values it can measure, the ranges it can be adjusted within, and the safety limits it must not cross are all recorded as natural-language tags. Reading the reference file built from this information lets even a brand-new agent understand a device's characteristics and allowed operating range.

There are three channels for controlling a connected device.

1. The Model Context Protocol (MCP), for agent connections.
2. A command-line interface (CLI), for direct human control.
3. Code files and APIs, for running fast or long-running tasks deterministically.

Elizabeth Kelly, Anthropic's head of responsible deployment, compared MHS to USB-C in a CNBC interview — a standard way to move information between devices. Just as software agents use MCP to connect to various tools, MHS extends that same standardization to physical equipment.

The second official video shows an AI model running an actual physical science experiment. Claude adjusted a laser, checked the beam's path with a camera, and adjusted again. It packaged what it learned through this iterative process into a deterministic code file, so that afterward it could run the alignment task with a single command instead of reasoning through every step again.

The numbers came from the lab first

Anthropic tested MHS first in genomics, robotics, and quantum computing research settings.

1. Genentech piloted automating BCA protein assays that combine a liquid handler, a robotic arm, and a plate reader.
2. The University of Washington's Baker and Fingleye labs built an agent that watches qPCR curves to decide the right moment to stop an experiment, along with collision-free plate handoffs between a robotic arm and a liquid handler.
3. Researchers at Carnegie Mellon University linked four incompatible pieces of equipment and three computers, running dose-response experiments roughly three times faster than before.
4. The HHMI Janelia research campus unified seven vendor programs that previously had no common interface. Anthropic's official announcement post said this cut the time needed for imaging experiments from weeks to a single day.
5. Quantum computing company QuEra built a controller where AI recovers laser lock whenever the laser drifts off its target frequency, recovering without human intervention 99.3% of the time.

Several manufacturing names are involved too. AWS supports MHS through Strands Robots, its library for connecting physical equipment, and plans to give research preview participants early access to a private pre-release version. Doosan Robotics is testing automated quality assurance for robotic arms and coordination of tasks across multiple robots. QIAGEN, Tecan, and Universal Robots are each building MHS support into their own equipment and platforms.

Not a sudden pivot

Back in July, Anthropic announced a physical AI partnership with UST to bring Claude into semiconductor verification and manufacturing processes. UST said its existing closed-loop verification pipeline cut verification cycles by 50-70%, shrinking work that normally took four days down to 48 hours, and the company plans to train 20,000 employees worldwide on Claude.

Where that earlier effort focused on embedding Claude into one company's production system, MHS is an attempt to build a shared interface across multiple manufacturers and multiple models. The starting point was a shared memory dictionary that HHMI Janelia postdoctoral researcher Arko Bast built to connect brain-imaging equipment combining lasers, motorized focusers, and cameras. Anthropic's Alec Kemeny and Bast then connected AI models to it, and it grew into a standard from there.

Still a limited research preview

MHS is currently application-only, available to a limited set of labs and manufacturers. Anthropic says it plans to build out safety evaluations and operational best practices during the research preview before releasing it as open source. It currently can't be used with equipment that lacks a programming interface.

A bigger limitation is the AI's physical intuition. During the Genentech experiment, when bubbles formed in a protein sample, Claude treated it like a software error. A researcher had to point out that bubbles can only be fixed through physical intervention. Anthropic says large language models, trained on text and images, still have incomplete spatial and physical reasoning — which is why expert oversight remains necessary.

Editor's take

The notable thing about this announcement isn't just that Claude moved a robotic arm. The bigger shift is that Anthropic isn't leaving decisions about what AI is allowed to do purely to the model's own judgment — it's trying to encode those limits into device driver reference files and safety boundaries instead. It's an attempt to build a standard boundary between AI, which reasons probabilistically, and physical systems, which have to move deterministically.

If integration time drops from months to minutes, automation projects that were once abandoned because of custom development costs could become realistic options. But the flip side is that a wrong command no longer just produces an on-screen error — it can damage a sample, crash equipment, or halt production. MHS's success will hinge less on how many devices it can connect and more on how consistently its safety limits hold up in real-world use. That's exactly what the research preview is meant to test.

Applications to join the MHS research preview are open at the official site.