MHS (Model Hardware Standard)

Connecting an AI agent — or any software — to a lab instrument or a factory machine normally means writing custom control code for that specific device, its firmware, and its vendor SDK. Per instrument, that work commonly runs from weeks to months. MHS is an attempt to standardize that layer the way MCP standardized how a model reaches software tools.

Anthropic previewed MHS in 2026 as a research preview, initially open to a limited group of scientific research labs and advanced manufacturers, and has said it intends to open-source the specification later. The work began as a collaboration with HHMI's Janelia Research Campus, with early case studies from labs and companies including Genentech, the University of Washington, Carnegie Mellon, and QuEra Computing. At the preview stage it is early, not a broadly adopted standard.

MHS standardizes the driver — the software layer between an operating system and a device. Each MHS driver exposes a small set of primitives: read (get a value, such as temperature), write (set a value), and discovery, so agents and devices can find each other on a network without a per-pair translator. The driver also carries natural-language tags describing what the device is, what it measures, which parameters are adjustable, and what safety limits it enforces; it compiles those tags into a reference file an agent reads to understand a device before operating it.

How it works

An MHS driver publishes its device on the network in a standard discovery format, so an agent can enumerate the available instruments without prior configuration. For each device the driver exposes the read/write primitives plus the compiled reference file — measurement units, adjustable parameters, and the safety limits the driver itself enforces, for example capping laser power so a sample isn't damaged. Because those limits live in the driver as code rather than in a prompt, they apply no matter which model is driving. The agent then operates the device through whichever control path its harness uses — MCP, a command-line interface, or generated code files — and can coordinate several instruments in parallel, since every driver speaks the same primitives. MHS is designed to be model-agnostic; Claude is Anthropic's reference implementation, not a requirement.

Example

In a case study Anthropic published, researchers at Carnegie Mellon connected several previously incompatible instruments spread across separate computers and went from writing a new driver to validated dose-response results in about eight hours — a workflow that vendor-standard integration would normally measure in weeks.

How it differs

Common misconceptions

FAQ