Anthropic previews a Model Hardware Standard for physical agents

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a model-agnostic standard that lets AI agents operate programmable microscopes, liquid handlers, robotic arms and other instruments. It was announced on August 27, 2026.

The standard can be accessed through standard protocols such as MCP, and it is being tested with selected laboratories and manufacturers ahead of an open-source release.

Why it matters: this is an attempt to create an interoperability layer between agent orchestration and many kinds of physical equipment. A common interface, however, is not a safety case.

Deterministic interlocks, calibrated state, bounded authority, signed identities, human approvals and safe-stop paths are still needed outside the model. Integration-speed claims are company-reported.