Anthropic previews AI hardware standard after early lab tests cut integration time and automate experiments

The Model Hardware Standard is being tested across scientific research and advanced manufacturing, with early projects spanning drug discovery, brain imaging and quantum computing

Anthropic has opened a research preview of its Model Hardware Standard, allowing AI agents to operate physical lab and manufacturing equipment

Anthropic has opened the first research preview of its Model Hardware Standard, a new specification designed to let AI agents operate physical equipment across laboratories and advanced manufacturing without building a separate integration for every device.

The company says MHS can reduce hardware integration from weeks or months to hours or minutes, while giving AI agents a common way to discover equipment, understand operating limits and coordinate multiple devices.

Early tests have already moved beyond simulations. Anthropic says AI agents using MHS have run a drug-discovery experiment with real-time error handling at Genentech, compressed an imaging experiment at HHMI Janelia Research Campus from weeks to a single day, and improved laser stabilization on QuEra quantum computers from 58% to 99.3%.

The system is not yet open source. Anthropic is first making it available to a group of scientific research labs and manufacturers so it can expand safety evaluations and gather evidence on how AI behaves when it is given direct control over physical equipment.

From bespoke integrations to a common hardware layer

The problem MHS is trying to solve is straightforward: laboratory and manufacturing equipment often comes with different programming interfaces, software and data formats.

Connecting devices together can therefore require specialists to build custom software for each setup. Adding an AI agent creates another layer of integration.

MHS introduces a standardized driver that translates between software and the physical device. It uses basic commands such as reading a temperature or changing a setting, while also describing the characteristics and safety limits of each machine.

That information can include details that might not be obvious from software alone, such as the weight of a robotic arm or the range within which a particular parameter can safely operate.

Once connected, an AI agent can receive data from several instruments, sequence their work, monitor outcomes and change parameters as conditions change.

Anthropic says MHS is model-agnostic and can work with any device that has a programmable interface. Agents can access hardware through mechanisms including the Model Context Protocol, command-line tools and APIs.

The standard began through a collaboration between Anthropic’s Beneficial Deployments team and HHMI Janelia Research Campus, where researcher Arco Bast was working with a brain-imaging setup combining lasers, cameras and motorized equipment from different manufacturers.

That collaboration has since expanded into robotics, electronics, biotechnology and manufacturing.

Genentech test exposed both the promise and the limits

One of the clearest early examples comes from Genentech, where researchers tested MHS as a proof of concept for automating a BCA protein assay.

The experiment required a liquid handler, robotic arm and microplate reader to work together. Claude acted as the central coordinator, with MHS providing the interface between the AI and each piece of hardware.

Researchers first asked Claude to run the assay using a standard protocol.

It did not get everything right. Claude initially selected generic liquid-handling settings for both water and a more viscous protein solution. That caused bubbles to form and produced inaccurate transfers.

The researchers then asked Claude to run experiments and optimize the flow rate itself. It tested different settings, analyzed readings from the plate reader and compared its work with an expert-performed transfer.

It eventually selected a flow rate of around 140 microliters per second for water and 10 microliters per second for the protein sample. Genentech says its automation specialists judged those settings reasonable for the setup.

The experiment also provided a useful reminder of what current AI systems still struggle with.

Claude recovered independently from some equipment errors, including failures involving pipette tips and fluid detection. But when bubbles caused problems during mixing, its initial response was to retry the operation with different parameters in the same well.

That made the problem worse. Researchers had to explain that the error came from the physical behavior of the liquid rather than a software fault, and tell the system to move to a clean well and reduce mixing.

Anthropic is using those kinds of failures to justify keeping MHS in a research preview before making it openly available.

The company acknowledges that large language models learn about the physical world mainly through text and images and can lack the spatial and physical intuition required for some real-world tasks.

University lab tests move AI closer to the bench

Academic researchers are also testing what happens when an AI agent can do more than analyze experimental data after the fact.

At the University of Washington, PhD student Zihao Song used MHS in the Baker and Pinglay labs to connect equipment used in protein research.

His experiments included a dashboard that allowed instruments to be monitored remotely, an AI-supervised quantitative PCR workflow and coordination between a robotic arm and liquid handler.

In the qPCR test, the system monitored amplification curves as they were produced and could halt the reaction when instructed before moving the equipment into a holding stage.

Another demonstration used Claude Code to coordinate a robotic arm and liquid handler during a plate handoff. Across repeated tests described by Song, the devices did not collide. The arm waited for liquid handling to finish before moving, while the liquid handler did not restart until the arm had cleared the plate.

Song also reported connecting six instruments through MHS in under a week, including the time required to write drivers for them.

Those demonstrations remain proofs of concept. He notes that more complicated laboratory protocols will require further optimization and that running AI agents continuously over long experiments also introduces compute costs.

That is part of the reason Anthropic is stopping short of presenting MHS as a finished automation layer.

The company is currently inviting researchers, manufacturers and other industry stakeholders to join the preview. Hardware companies including Automata, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan and Universal Robots are among those testing or building support for the standard, while AWS is supporting the preview through Strands Robots.

Hugging Face is also adding MHS support to LeRobot, and Raspberry Pi is enabling integration across a number of its products following tests with a Camera MHS Driver.

Anthropic says it will use the research preview to develop additional physical safety evaluations and publish guidance when MHS is eventually released as open source.