Anthropic Model Hardware Standard (MHS) research preview

TL;DR

Anthropic announced a research preview of the Model Hardware Standard (MHS), a shared driver and protocol that lets AI agents communicate with any programmable instrument, cutting integration from weeks to minutes and enabling round‑the‑clock autonomous experiments.

What is the Model Hardware Standard?

MHS is a model‑agnostic, device‑agnostic specification that defines a minimal set of primitives (e.g., read, write) and a discovery format so that any hardware with a programmable interface can be accessed by an AI agent through a single, standard driver. It was co‑developed with HHMI Janelia Research Campus and is built on top of the existing Model Context Protocol (MCP).

Key properties:

Standard driver translates OS calls to device‑specific commands.
Natural‑language tags allow users or agents to annotate physical characteristics (e.g., robot arm weight) that are not exposed by code.
Safety limits are enforced by the driver, preventing agents from issuing dangerous commands.
Model‑agnostic: any Claude or other LLM can use MHS via the same MCP interface.

How MHS Works

Device discovery – Each instrument publishes a JSON‑like descriptor that lists its measurable states, controllable parameters, and safety constraints.
Primitive commands – Agents issue simple read/write calls; the driver maps them to the device’s native API (CLI, SDK, or GUI automation).
State dictionary – All device states are stored in a shared‑memory dictionary, readable by any process, enabling real‑time monitoring and cross‑device analytics.
Orchestration layers – Agents can chain commands in code files (APIs) for high‑frequency operations, or use MCP for interactive control.
Error handling – The driver validates each command against safety tags and can automatically pause or abort if a violation is detected.

Early Partner Demonstrations

Genentech – BCA protein assay automation

Integrated a liquid handler, robotic arm, and plate reader.
Claude optimized flow rates for water (≈ 140 µL/s) and viscous BSA (≈ 10 µL/s) using closed‑loop RMSE feedback.
Demonstrated autonomous error recovery from tip‑pickup and foaming failures.

University of Washington – Remote monitoring & qPCR control

Dashboard aggregated status of all instruments via MHS.
AI‑supervised qPCR halted runs at the optimal amplification curve, reducing manual watch‑time.
Coordinated plate handoffs between a liquid handler and a robotic arm without collisions.

Carnegie Mellon University – Rapid dose‑response curves

Combined a CyBio Felix liquid handler, Varioskan LUX plate reader, robotic arm, and cameras.
MHS reduced integration time to ~8 hours and accelerated serial‑dilution experiments three‑fold.
Autonomous safety checks blocked six simulated fault conditions before any motion.

HHMI Janelia – Microscopy rig unification

Unified seven vendor programs into a single MHS dictionary.
Added a new camera in minutes; the camera’s beam‑position data fed directly to galvanometer mirrors for sub‑micron alignment.
Enabled online analysis pipelines that read data from the shared dictionary regardless of the originating language (MATLAB, Python, C#).

QuEra Computing – Quantum laser stabilization

MHS gave Claude control over a titanium‑sapphire laser lock.
An iterative agent loop rewrote the recovery script, cutting lock‑recovery time from ~150 s (58 % success) to ~6 s (96 % success) and ultimately achieving 99.3 % success in blind tests.
Claude also tuned 12 PID parameters, reducing RMS noise ten‑fold and eliminating lock drops over a 19‑hour run.

Tetsuwan Scientific – qPCR for citizen‑science water monitoring

Integrated cameras, centrifuges, and liquid handlers through MHS.
Claude detected bubbles via computer‑vision, instructed a centrifuge to spin down the tube, and resumed the protocol automatically.
Closed‑loop compiler optimization improved multi‑dispense precision by ~12 % (p ≈ 0.001).

Safety, Limitations, and Ongoing Work

Physical reasoning gaps: Claude still misinterprets phenomena such as bubble formation; human guidance is required to teach correct handling strategies.
Interface requirement: MHS only works with devices that expose a programmable interface (API, CLI, or GUI). Non‑programmable hardware is being retrofitted with custom drivers.
Human‑in‑the‑loop: For high‑risk actions, Claude pauses for human approval; future model upgrades aim to reduce unnecessary pauses while preserving safety.
Safety roadmap: Anthropic is drafting a physical‑safety policy, expanding fault‑injection tests, and planning formal safety‑evaluation benchmarks before open‑sourcing the standard.

Future Roadmap and Open‑Source Plans

Open‑source release – After completing safety evaluations and expanding driver coverage, MHS will be published under an open‑source license.
Broader device support – Adding centrifuges, incubators, qPCR thermocyclers, and other non‑API equipment.
Standard driver library – Public repository of ready‑made drivers for common lab and manufacturing instruments.
Community sandbox – A hosted testbed where researchers can submit new drivers, share safety cases, and benchmark AI‑driven workflows.
Integration with other Anthropic tools – Tight coupling with Claude Code, Claude Opus, and the Model Context Protocol for seamless end‑to‑end pipelines.

How to Join the Research Preview

Interested labs or manufacturers can apply via the MHS portal: https://www.modelhardwarestandard.com/.
Selected participants receive a private pre‑release of the Strands Robots library (AWS) and early driver builds.
Participants are expected to co‑author safety evaluations and share integration experiences with the Anthropic team.

Acknowledgments

The standard originated from a collaboration between Alek Kemeny (Anthropic Beneficial Deployments) and Arco Bast (HHMI Janelia). Additional contributors include Aaron Boswell, Ben Arthur, Boaz Mohar, Gagan Bhat, Mark Kittisopikul, Nadine Yasser, Nick Purcell, Takashi Kawase, and Virginie Ruetten.

Conclusion

MHS demonstrates that a unified hardware abstraction layer can shrink integration cycles from weeks to minutes, enable AI agents to run complex, multi‑instrument experiments autonomously, and provide a safety‑first foundation for future AI‑driven laboratories and factories. The research preview invites the community to help validate, extend, and harden the standard before it becomes an open‑source cornerstone for AI‑instrument interaction.