Artificial intelligence is given control of laboratory robots and microscopes for the first time

Anthropic recently made available the Anthropic Model Hardware Standard (MHS), a common specification that enables AI agents to securely control physical devices in the real world. Access is currently limited to a select group of research labs and advanced manufacturers, with a waiting list open to other interested parties. Anthropic says the Anthropic Model Hardware Standard (MHS) reduces the integration time of complex hardware from weeks or months to hours or even minutes.

The new standard is designed for microscopes, automated pipetting systems and robotic arms, with a single agent being able to control multiple devices in parallel. Anthropic lists use cases ranging from routine drug discovery experiments to precise alignment of laser beams in quantum computers. The goal of the project is to have fully autonomous experiments that run 24/7, with the agent analyzing each step, adjusting parameters as conditions change, and in some cases even recovering from a hardware failure without human intervention. Jonah Cool, from Anthropic’s biology department, says the standard solves a key problem: a lot of science isn’t being done today simply because researchers can’t properly operate the equipment they have.

The development of the Anthropic Model Hardware Standard (MHS) is the result of a collaboration between Anthropic and HHMI's Janelia Research Campus. The idea was born from observations by neuroscientist Arca Basto in a memory formation experiment, for which he himself designed a common interface to coordinate rotating laser beams, microscopes, and cameras. The idea was developed into a universal protocol for autonomously running scientific experiments, which functions similarly to Anthropic's existing Anthropic Model Context Protocol (MCP) software protocol.

At the heart of the Anthropic Model Hardware Standard (MHS) is a standardized driver that translates commands between the operating system and the attached device. The driver uses a modest set of basic commands, such as reading data or setting parameters, that any instrument can receive. In addition, the system includes security constraints and device specifications, which are written directly in natural language text tags.

The standard has already been tested in Genentech’s laboratories on an automated protein concentration station, where the Anthropic Claude model controlled a pipettor, robotic arm, and microplate reader. After initial adjustments to the fluid flow, the model independently optimized parameters for water and protein that would otherwise require manual programming by an expert. Partners currently testing the Anthropic Model Hardware Standard (MHS) include Amazon Web Services (AWS), Hugging Face and Raspberry Pi Ltd. At the end of the preview phase, Anthropic plans to offer the standard as an open source solution.