Anthropic pilots MHS, an MCP-style hardware protocol letting Claude control robots and lab tools

Anthropic staff dub the new protocol MHS and describe it explicitly as the hardware counterpart to the Model Context Protocol (MCP) that already connects Claude and ChatGPT to apps like Gmail, Google Calendar, and Slack. Where MCP standardizes how a model calls software tools, MHS aims to standardize how a model discovers and drives physical hardware—robots, lab instruments, and manufacturing equipment—so Claude can plan and execute real experiments rather than just describe them.

Mechanically, the research preview lets Claude read a hardware device's capabilities and issue commands to run multi-step procedures autonomously, with safeguards layered in given the obvious physical-world risk. Anthropic's Alek Kemeny stressed the scientific angle: automating experiment planning, protocol generation, execution, and analysis in a lab loop. Partners named in testing include AWS and Hugging Face.

Competitively, this lands the same week Google DeepMind detailed its AI Co-Scientist running a full closed research loop—planning experiments, writing code, controlling lab equipment, and drafting manuscripts. Both signal a race to move agents off the screen and into wet labs and robotics, where the value (and the safety stakes) are far higher. NVIDIA's Cosmos 3 world models for physical AI point at the same frontier from the simulation side.

The caveat: a hardware-control standard from a single lab invites the same fragmentation and safety questions MCP raised, amplified by physical consequences—an errant command in a bio or chemistry lab is not a reversible software error. Reproducibility and access control will be scrutinized. Watch whether MHS becomes an open standard others adopt, or stays an Anthropic-controlled preview.