Anthropic opens Model Hardware Standard research preview

• Anthropic has opened MHS as a research preview for scientific labs and advanced manufacturers, where AnthropicAI’s announcement describes a shared specification for agents to operate physical equipment.
• MHS replaces bespoke hardware bridges with a standard driver, and AnthropicAI says integration can fall from days or weeks to hours or minutes.
• Early deployments span a Genentech drug-discovery experiment, HHMI microscopy, and QuEra quantum hardware, with AnthropicAI’s pilot list reporting laser stabilization rising from 58% to 99.3%.
• The preview remains deliberately constrained: AnthropicAI says language models lack physical intuition and still require expert oversight.

The system exposes three control paths, while AnthropicAI’s announcement says device traits can become natural-language tags and an auto-generated reference file. An Ars Technica report captures the immediate use case: replacing the custom integration work needed to make mismatched lab equipment cooperate.

Anthropic’s launch post defines MHS as a software adapter between the operating system and a physical device. Its core pieces are:

• Read and write primitives: commands such as getting or setting temperature.
• Standard discovery: agents and devices can locate one another across a network without a one-off translator.
• Natural-language tags: operators can record characteristics that code may omit, including a robot arm’s weight.
• Reference files: tags generate a device description covering measurements, adjustable parameters, and enforced safety limits.

The setup can also have an agent interview an operator about the hardware before producing those tags.

After a device is described and connected, MHS offers three ways to control it, according to the technical overview:

• MCP, for agent-harness access through a standard protocol.
• A command-line interface, for direct device operations.
• Code files (APIs), for composing driver commands into a reusable program.

Anthropic says code files let hardware execute long-running or fast operations without requiring online model reasoning at every step. The company observed Claude adjusting a laser, inspecting camera feedback, and then packaging the learned procedure into a deterministic alignment script that runs as one command.

Five of the six examples in Anthropic’s launch post show the standard being used as an orchestration layer rather than a single-device remote control:

• Genentech: a BCA protein assay coordinated a liquid handler, robotic arm, and plate reader.
• University of Washington: a remotely monitored setup included an agent-supervised qPCR that watches amplification curves and stops at the appropriate point, plus collision-free plate handoffs between a robot arm and liquid handler.
• Carnegie Mellon: serial-dilution dose-response experiments ran about three times faster across a liquid handler, plate reader, robotic arm, and cameras spread across three incompatible computers.
• HHMI Janelia: a microscopy rig previously split across seven vendor programs gained a shared interface.
• Tetsuwan Scientific: its ResearchOS platform used MHS to orchestrate qPCR work for a San Pedro Creek pollution-monitoring project.

According to imjustnewatai’s QuEra recount, four specialists had spent months producing a laser-lock recovery script that succeeded 58% of the time and took about 150 seconds per attempt. The post says four agents then revised and tested the script against a live laser overnight, reaching 96% recovery in six seconds before a blind test recovered 695 of 700 attempts.

Anthropic’s partner description gives the public result as a controller that restored the lock, the frequency precision needed to interact with neutral atoms, 99.3% of the time without human intervention.

An agent calibrates a laser through MHS

Anthropic’s launch post says Claude learns the physical world from text and images, leaving spatial and physical reasoning gaps that require expert oversight. At Genentech, researchers had to teach it that foaming in protein samples was a physical failure rather than a software bug, and that correction required a physical intervention.

The research preview is also where Anthropic plans to build additional physical-safety evaluations with launch partners, develop deployment practices, and expand safeguards-policy coverage for misuse risks.

The standard currently requires a programmable device interface, and Anthropic gives no open-source date. The Register’s report notes that the protocol itself is not yet public.

The launch post names AWS, which will provide preview participants a private pre-release Strands Robots package; Automata, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, and Universal Robots as hardware and automation partners. Hugging Face is adding MHS support to LeRobot, while Raspberry Pi is extending an initially successful Camera MHS Driver across more products.