Anthropic Opens Research Access to the Model Hardware Standard for AI Agents and Laboratory Instruments

Anthropic Opens Research Access to the Model Hardware Standard for AI Agents and Laboratory Instruments

On August 27, Anthropic launched research access to the Model Hardware Standard (MHS). The standard gives AI agents a common way to retrieve data from instruments, change permitted settings, and operate within defined safety limits. Early partner pilots have already used MHS to coordinate several laboratory devices.

A laboratory experiment can be delayed long before the first measurement. Cameras, robotic arms, and sample readers often use different software, so each new combination requires a separate interface. According to Anthropic's announcement, configuring and integrating equipment usually takes weeks and sometimes months.

MHS provides a standard driver for each instrument. A driver is a program that connects the operating system to the hardware. A read command can send the agent a temperature measurement, for example, while a write command changes a setting. The user specifies what the instrument measures, which parameters can be changed, and which safety limits apply. The driver compiles this information into a reference file for the agent.

This allows the agent to monitor the experimental setup and coordinate steps across several instruments. It can retrieve a reading, change a setting, and verify the result. Commands for a repeated operation can be combined into a fixed procedure, which is code that executes an established sequence without requiring the agent to reason through each step again.

In June, Janelia described a cycle in which a model proposes the next experiment after receiving new measurements. MHS gives such a cycle a common way to retrieve readings and send permitted commands to instruments.

The project grew out of a problem faced by Arko Basta, a postdoctoral researcher at the Janelia Research Campus. His research microscope combined cameras, scanners, and sensors that used different control systems, so preparing a new experimental setup took months. An HHMI article describes Basta's solution: an application that allows every component of the setup to read and update a shared representation of its state. According to Janelia, this reduced experiment preparation time from months to days. In February 2026, Basta presented his work to Anthropic engineers. MHS grew out of the collaboration that followed.

“Setting up an experiment takes much less time. That gives me more time to think about the science,” Basta says.

In a Carnegie Mellon University pilot, researchers automated serial dilutions. This process dilutes a solution step by step to construct a dose-response curve, which shows how the result depends on concentration. The agent coordinated a liquid-handling robot, an instrument that read samples from a multiwell plate, a robotic arm, and cameras across three computers with incompatible interfaces. The test used a dye instead of a drug candidate. Anthropic reports that MHS helped complete the series about three times faster.

Research access to MHS is available by application. Before the public release, Anthropic is working with laboratories and instrument manufacturers to develop safety checks and rules governing how AI interacts with physical equipment.