MCP connected AI to software. MHS connects AI to the physical world.

MCP connected AI to software. MHS - Anthropic's Model Hardware Standard - connects AI to the physical world.

Just like the Model Context Protocol gave agents a universal standard to talk to APIs and databases, MHS is a unified driver layer for agents to control physical hardware: microscopes, robotic arms, liquid handlers, lasers, sensors.

The problem it solves

Previously, every lab device had its own closed SDK. Connecting seven devices took weeks of custom glue code.

MHS introduces universal read/write primitives, hardware capability tags, and safety envelopes. Plug a device in, the agent discovers its capabilities and controls it deterministically or dynamically.

The "USB / MCP moment" for physical machines and labs.

Real examples from the research preview

• Genentech - Claude read a research paper PDF, commanded a robotic arm and liquid handler, executed the biology experiment overnight, and self-corrected glitches on its own.
• QuEra Quantum Computing - Claude controlled lasers inside a neutral-atom quantum computer, restoring laser lock 99.3% of the time (695/700) vs manual human calibration.
• HHMI Janelia - Claude autonomously drove a Leica microscope, auto-focusing on neurons and hunting for bacteria samples live.

The safety part I love

MHS drivers enforce strict physical boundaries - robotic arm travel ranges, preventing the microscope objective from crashing into glass slides. The AI physically cannot break the hardware.

What it means

Scientists can drop an experimental protocol PDF into Claude and let the agent run and iterate experiments autonomously. The same universal standard MCP gave software is now giving hardware - and that is a genuinely big export.