Anthropic’s MHS: the new interface between AI agents and machines

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices. The first focus is scientific research and advanced manufacturing, with a planned path toward open source. 1

Instead of teaching one agent one instrument, MHS gives devices a common driver layer: simple read/write commands, metadata about ranges and safety limits, and ways to connect through MCP, command line, or code. Anthropic says integration work that once took weeks or months could shrink to hours or minutes. 12

Early partner examples make the pitch concrete: Carnegie Mellon reported about 3× faster serial-dilution dose-response experiments; QuEra used the standard for laser-lock recovery 99.3% of the time without human intervention; Genentech coordinated a liquid handler, robotic arm, and plate reader in a protein assay. 1

Why it matters: agent progress can be measured in completed physical workflows, not only text or code. The boundary is equally important: MHS currently depends on programmable hardware, selected-partner testing, and expert oversight for physical, chemical, and biological edge cases. 23