What Is Anthropic's Model Hardware Standard (MHS)?
The Short Answer
MHS is MCP for machines.
Anthropic announced the Model Hardware Standard on August 27, 2026 as a research preview. It is a common description format that lets a piece of physical equipment tell an AI agent three things: what it can do, what it can measure, and what it must never be asked to do.
That third part is the interesting one. Every previous attempt at “let the model drive the robot” put the safety envelope in the prompt or in bespoke glue code. MHS puts it in the device’s own specification.
The Problem It Solves
Integrating a lab or factory device with software today means custom work — often weeks or months per device — because the knowledge of how the machine actually operates lives in PDF manuals and in the heads of a few experts.
A science lab may run dozens or hundreds of instruments. Anthropic’s Alek Kemeny, a member of technical staff, put the value plainly: “Where it provides value is in science labs, where you have dozens or hundreds of devices.”
That is the scaling argument. One integration is a project. Three hundred integrations is a standard.
What an MHS Specification Contains
| Layer | What the device declares | Why it matters |
|---|---|---|
| Capabilities | operations the machine can perform | agent discovers actions without a manual |
| Measurements | what it senses and reports | agent can close the loop on results |
| Settings | adjustable parameters and ranges | agent can configure, not just trigger |
| Safety envelope | hard limits — speed, angle, force, temperature | constraint enforced at the device, not the prompt |
Anthropic’s own worked example: a factory robot arm vendor specifies in its MHS how AI may manipulate a heavy arm safely by limiting the speed or angles through which it moves. The vendor — who understands the machine — owns the constraint. The agent inherits it.
Who Is Testing It
Named early partners as of the August 27, 2026 announcement:
- Amazon Web Services
- Danaher (life sciences and diagnostics instruments)
- Hugging Face
- Raspberry Pi
The mix is telling: one hyperscaler, one instrument manufacturer, one ML platform, one hobbyist-to-industrial hardware vendor. Anthropic is testing whether the standard generalises across the whole range rather than fitting one vertical.
The showcase demonstration came from Genentech, where a scientist sent Claude a PDF of an experiment they had designed and the assistant executed it autonomously on MHS-equipped hardware. Jonah Cool, who leads partnerships and deployment at Anthropic’s life sciences arm, framed the motivation as access rather than speed: “In many cases, the science doesn’t happen because you can’t use the equipment or it’s too technically challenging.”
Status and Availability
| Detail | |
|---|---|
| Announced | August 27, 2026 |
| Stage | Research preview |
| Access | Waitlist, selected scientific / robotics / manufacturing orgs |
| Open source | Intended after preview — no timeline given |
| Safety work | Evaluations and secure-deployment guidance promised before broad release |
Anthropic is explicitly sequencing safety before availability: the preview exists so developers can “test how Claude interacts with the physical world and build in necessary safeguards before releasing it broadly.”
Why This Is a Bigger Deal Than It Looks
Three reasons to take MHS seriously rather than filing it as a robotics demo:
- Anthropic has done this before and it worked. MCP went from an Anthropic proposal to a de facto industry standard adopted well beyond Claude. The company has demonstrated it can ship a protocol the rest of the market picks up.
- The market is contested. Google, OpenAI and Nvidia are all developing models and software for robotics. A Barclays report earlier in 2026 projected AI-powered robots and autonomous machines as a trillion-dollar market by 2035. Standards get set early in contested markets.
- The failure mode is physical. A software agent that misreads an API returns a bad result. A hardware agent that misreads a device breaks something or hurts someone. Putting the safety envelope in a machine-readable spec is the structurally right answer, and it is a meaningfully different design problem from MCP.
What to Do About It Now
If you build or operate lab, robotics or manufacturing equipment: join the waitlist and read the spec when it lands. Early participation in a standard that may become default is cheap; retrofitting to it later is not.
If you build agent software: nothing to integrate yet, but note the pattern. Declarative, device-owned safety envelopes are a design worth copying even before MHS is open — it is a better answer than trusting a system prompt to hold a limit.