Anthropic Previews MHS to Standardize AI Control of Lab and Factory Hardware
Anthropic previews MHS, an AI hardware standard cutting lab integration from months to minutes, after early tests at Genentech, UW and Carnegie Mellon.
Summary
On August 26, 2026, Anthropic opened the Model Hardware Standard research preview to an initial group of scientific laboratories and advanced manufacturers. Developed with HHMI Janelia Research Campus, MHS lets model-agnostic AI agents safely coordinate programmable microscopes, liquid handlers, robotic arms and other equipment in parallel, potentially reducing bespoke integrations from weeks or months to hours or minutes. Partners will develop safety evaluations and best practices before the standard becomes open source.
MHS drivers translate device interfaces into common read and write commands, advertise equipment across networks and use natural language tags to document characteristics, capabilities and enforced safety limits. Agents control devices through Model Context Protocol, command line tools or APIs, sequence workflows, adjust parameters from live data and package time-sensitive or long-running operations into deterministic code. Claude demonstrated this by repeatedly adjusting and observing a laser, then writing a single-command alignment script.
Genentech used Claude to orchestrate a liquid handler, robotic arm and plate reader for a 96-well BCA protein assay. It optimized water to about 140 µL/s with 0.016 RMSE and viscous BSA to 10 µL/s with 0.181 RMSE, but needed human guidance when retries worsened bubble-related failures. University of Washington PhD student Zihao Song connected six instruments in under a week, automated qPCR stopping and a 4 °C hold, and completed repeated LeRobot plate handoffs without collisions, triggering arm movement about 10 seconds after dispensing. Carnegie Mellon University ran serial dilution dose response experiments about three times faster across four device types and three computers with incompatible interfaces. More complex protocols still require optimization, additional hardware integration and assessment of continuous-agent compute costs.
Positives
- MHS reduced hardware integration from weeks or months to hours or minutes by replacing bespoke connections with standardized drivers.
- Genentech’s Claude workflow identified reasonable flow rates of about 140 µL/s for water and 10 µL/s for viscous BSA.
- Zihao Song connected six University of Washington instruments through MHS in under a week, including driver development.
- Repeated LeRobot plate handoffs completed without collisions, with arm movement beginning about 10 seconds after liquid dispensing ended.
- Carnegie Mellon completed serial dilution dose response experiments about three times faster across incompatible devices and computers.
- MHS supports any programmable device, remains model-agnostic and exposes control through MCP, command line tools and APIs.
Risks & concerns
- Claude repeatedly agitated bubble-filled wells until Genentech researchers explained the physical cause and directed it toward gentler handling.
- More complicated experimental protocols need substantial optimization and broader physical manipulation capabilities before reliable autonomous operation.
- Continuous AI monitoring over long experiments creates compute costs that must be weighed against researchers’ saved time.
- MHS remains an early research preview, with safety evaluations and operating best practices still required before open sourcing.
- Genentech must add centrifuges, incubators, analytical instruments and sensors before MHS can support broader drug discovery workflows.
