Anthropic wants AI models to control physical hardware

Anthropic is proposing an open standard that lets AI agents directly control physical hardware, from lab microscopes to industrial robot arms.

Ars Technica reports that Anthropic announced the Model Hardware Standard (MHS). The system uses standardized reference files to tell AI models about a device's physical limits, safety bounds, and controls. Instead of retraining a model for every tool, MHS lets it write API scripts to run instruments it has never seen before. In one demo, Claude reasoned out how to pick up an aluminum can with a robot arm without step-by-step training. Anthropic is previewing the tech with AWS, Hugging Face, Raspberry Pi, Automata, and Universal Robots.

Why it matters: Connecting AI to physical hardware usually requires custom code for every single instrument. Standardizing how models read device specs and send commands cuts setup times. Anthropic reported that early testing with scientific partners over the last year helped teams iterate faster on experiments.

Anthropic says MHS will eventually be an open-source, agent-agnostic standard, so it won't be locked to Claude.

If it catches on, hooking an AI up to a piece of lab equipment might soon feel more like plugging in a USB mouse.