Anthropic unveils MHS, a universal standard for AI agents that control real‑world machines

Anthropic, the company behind Claude, has announced a research preview of the Model Hardware Standard (MHS), a common interface for AI agents that operate real‑world machines such as microscopes and robotic arms. The standard is being co‑developed with the Howard Hughes Medical Institute (HHMI), which is already testing it on high‑end lab microscopes to automate complex imaging workflows.

MHS is explicitly model‑agnostic: any large language model can, in principle, use it to send commands to hardware and read back sensor data in a structured way. That could let labs and factories swap in different AI agents without rewriting low‑level control software for every vendor. At the same time, letting autonomous systems directly steer physical equipment raises fresh safety and liability questions. Anthropic says access will stay limited to selected partners while it works with them to build in guardrails, logging and human‑in‑the‑loop controls before opening the standard more broadly.