Anthropic's Model Hardware Standard lets agents safely operate lab equipment

APPSO · wechat · 2026-08-28

Anthropic has released a research preview of Model Hardware Standard (MHS) — instead of building AI a dedicated body, it lets agents temporarily inhabit and operate existing physical hardware. In a demo, Claude used MHS to coordinate a liquid handler, cameras, a robot arm, and a centrifuge from different vendors to autonomously resolve a pipetting bubble anomaly without interrupting the experiment.

MHS is not simply "MCP for hardware": MCP standardizes how AI calls software, while MHS standardizes how a machine describes itself to an agent — its capabilities, state, and safety limits. The control chain runs user goal → agent → MCP/CLI/API → MHS driver → device. Each device connects once and can then interoperate with all other MHS devices at bare-metal speed, potentially cutting multi-device integration from weeks to hours.

Early results are striking: CMU used MHS to orchestrate four device types across three computers, speeding experiments 3x; quantum computing firm QuEra let Claude recover laser lock faults on a $700K testbed — 695 of 700 deliberately induced failures fixed (99.3%), worst case 10-14 seconds. MHS is model-agnostic: any agent framework can connect. It remains an application-only preview, not open source, and Anthropic notes Claude's spatial/physical reasoning still requires expert supervision — in the physical world there is no Ctrl+Z.

Related event: Anthropic Launches Model Hardware Standard to Let AI Agents Safely Control Physical Equipment(25 posts)→

Original post →

More from coding & agent

coding & agent channel →