Anthropic unveils MHS, letting AI operate lab equipment

The first official video shows MHS linking different pieces of lab and factory equipment under a single agent. The AI doesn't stop at calling tools on a screen — it moves on to reading sensors and sending commands directly to robotic arms and lab instruments.

MHS is the USB-C of hardware

Lab equipment looks uniform from the outside, but each device speaks its own language. Microscopes, plate readers, and liquid handlers all have different programming interfaces, so linking several instruments into one experiment meant an expert had to build a custom translator every time. Even once the equipment was connected, there was no common way for an AI agent to read data and issue commands safely.

MHS inserts a standard driver in between. The driver provides basic commands every device can understand in common — things like read to pull a temperature reading or write to set one. A device's weight, the values it can measure, its adjustable ranges, and the safety limits it must never cross are all recorded as natural-language tags. Any agent encountering the equipment for the first time can read the resulting reference file and immediately understand the device's characteristics and allowed operating range.

There are three channels for controlling connected equipment.

1. Connect to agents through the Model Context Protocol (MCP).
2. Control equipment directly through a command-line interface (CLI).
3. Run fast or long-running tasks deterministically through code files and APIs.

Elizabeth Kelly, Anthropic's head of beneficial deployment, compared MHS to USB-C in a CNBC interview — a standard that governs how information passes between devices. Just as software agents used MCP to connect to a range of tools, MHS extends that same standardization approach out into physical equipment.

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

The numbers came out of the lab first

Anthropic tested MHS first in life-science, robotics, and quantum-computing research settings.

1. Genentech piloted automating BCA protein assays that combine liquid handlers, robotic arms, and plate readers.
2. The Baker and Pinglay labs at the University of Washington built an agent that watches qPCR curves to stop experiments at the right moment, along with collision-free plate handoffs between robotic arms and liquid handlers.
3. Researchers at Carnegie Mellon University linked four incompatible instruments and three computers together, running dose-response experiments roughly three times faster than before.
4. HHMI's Janelia Research Campus unified seven vendor programs that previously had no common interface. Anthropic's official announcement post says imaging experiments that used to take weeks now take a single day.
5. Quantum computing company QuEra built a controller in which AI restores laser lock whenever the frequency drifts off target. It recovered without human intervention 99.3% of the time.

A number of manufacturing names are involved too. AWS is supporting MHS through Strands Robots, its library for connecting physical equipment, and plans to give research-preview participants an early private version. Doosan Robotics is testing automated quality assurance for robotic arms and task coordination across multiple robots. QIAGEN, Tecan, and Universal Robots are also adding MHS support to their own equipment and platforms.

This isn't a sudden pivot

Back in July, Anthropic and UST announced a physical AI partnership that put Claude into semiconductor verification and manufacturing processes. UST said its existing closed verification pipeline cut verification cycles by 50-70%, shrinking work that typically took four days down to 48 hours, and committed to training 20,000 employees worldwide on Claude.

If that earlier move was about embedding Claude inside one company's production system, MHS is an attempt to build an interface that many manufacturers and many models can share. It started as a shared-memory dictionary that HHMI Janelia postdoc Arko Bast built to link brain-imaging equipment combining lasers, motorized focusers, and cameras. Anthropic's Alec Kemeny then connected an AI model to that dictionary alongside Bast, and it grew into a standard from there.

Still a limited research preview

MHS is currently accepting applications and is only available to a select group of labs and manufacturers. Anthropic says it plans to develop safety evaluations and operational best practices during the research preview before releasing MHS as open source. For now, it can't be used with equipment that lacks a programmable 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 bug. A researcher had to step in and explain that bubbles can only be fixed through a physical intervention. Anthropic's explanation is that large language models, having learned about the world through text and images, still reason imperfectly about space and physics — which is why expert oversight is still required.

Editor's view

What matters most in this announcement isn't simply that Claude moved a robotic arm. The bigger shift is that Anthropic isn't leaving decisions about what the AI is allowed to do purely to the model's own judgment — it's encoding those boundaries into the equipment driver's reference files and safety limits instead. It's an attempt to draw a standard boundary between an AI that reasons probabilistically and physical systems that need to move deterministically.

If integration time drops from months to minutes, automation projects that used to get shelved over custom development costs could suddenly become realistic. But the flip side is that a bad command no longer just shows up as an error on screen — it can damage samples, crash equipment, or halt production. Whether MHS succeeds will come down less to how many devices it can connect and more to how consistently its safety limits hold up in real-world use. That's exactly what the research preview needs to prove out.

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