Until now, getting an AI agent to operate a laboratory instrument meant weeks of custom engineering work before a single experiment could run. Anthropic's new Model Hardware Standard (MHS) is designed to change that, reducing the setup time to hours or minutes by giving agents a common language for talking to physical devices.
The company released MHS on 27 August 2026, initially as a research preview. Anthropic says development began as a collaboration between the company and the Howard Hughes Medical Institute (HHMI) Janelia Research Campus. Anthropic plans to open-source the standard after working with partners on safety evaluations, though no timeline for that release has been set.
The practical target is laboratories and manufacturing floors where several instruments must work in concert. A drug discovery workflow, for instance, might require a liquid handler, a robotic arm, and a plate reader to coordinate in sequence. Previously, each connection between those devices typically needed a bespoke translator program written by hand. According to Anthropic, MHS replaces that with standardised drivers and a common data format, so devices can share information across a network without custom glue code in between.
The specification works with any device that has a programmable interface. Its drivers use simple building blocks: a 'read' primitive, such as 'get temperature', and a 'write' primitive, such as 'set temperature'. Anthropic says any agent harness can access MHS using standard protocols, including the company's own Model Context Protocol (MCP), and that the standard is model-agnostic, meaning it is not restricted to Claude or any other particular AI system.
Anthropic tested the standard using Claude on a quantum computing task: aligning a laser. During that testing, Claude adjusted the laser, observed the results through a camera to assess how its adjustment moved the beam, and repeated the process. It then packaged what it had learned into code files, producing a deterministic script that allowed it to align the laser without reasoning through each step from scratch on future runs.
Genentech ran a separate proof-of-concept, using MHS to automate the BCA protein assay, a standard biochemical test. That required coordinating a liquid handler, a robotic arm, and a plate reader working together, according to Anthropic.
The intended audience is researchers and engineers who want to run autonomous, round-the-clock workflows with minimal human intervention. Anthropic positions MHS as a way to remove the bottleneck of custom software integration, which the company says has slowed the adoption of AI-driven automation in science and manufacturing. Before any device can be put under agent control, teams have typically had to write and maintain integration code for every instrument involved.
Anthropic says it is sharing an early version of the standard with partners across science, robotics, electronics, and manufacturing ahead of a broader open-source release. The safety evaluation process and criteria have not been made public, and it is not known how many or which organisations beyond Genentech and HHMI Janelia are currently testing MHS. Pricing and licensing terms for commercial use have also not been disclosed.
Comments (0)