Previewing the Model Hardware Standard

Summary

Anthropic and HHMI Janelia Research Campus are opening a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to operate programmable laboratory and manufacturing devices safely. MHS replaces bespoke integrations with standardized drivers built around read and write primitives, device discovery, natural-language tags, generated reference files, and control through MCP, a command-line interface, or code files. The model-agnostic standard lets agents coordinate instruments, monitor operating data, adjust parameters in real time, and turn exploratory work into deterministic scripts; early projects covered protein assays, qPCR, microscopy, robotics, dose-response experiments, and quantum-laser stabilization. Reported results include serial dilution experiments running about three times faster and a QuEra laser-lock controller recovering 99.3% of the time, while the preview will inform safety evaluations and guidance before an open-source release.

Takeaways

1. 1

   MHS drivers expose device operations through simple read and write primitives, make hardware discoverable across networks, and provide agents with generated reference files containing characteristics and enforced safety limits.

2. 2

   Carnegie Mellon researchers used MHS to run serial dilution dose-response experiments about three times faster, coordinating a liquid handler, plate reader, robotic arm, and monitoring cameras across three computers.

3. 3

   MHS currently requires a programmable hardware interface, and physical reasoning limitations mean expert oversight remains necessary; the research preview will be used to develop additional safety evaluations and protections.