Anthropic didn't enter robotics — it ran the MCP playbook on the physical layer

What actually shipped

A specification, a gated preview, and a partner list.

MHS standardises how a physical device describes itself to an AI agent. Underneath, it is deliberately unglamorous: a driver layer built on primitives as simple as read and write, plus natural-language tags that capture what a piece of hardware is and what it must never be asked to do. Device-level safety limits are enforced at that layer rather than trusted to the model. Anthropic’s claim is that integrating an instrument drops from weeks or months to hours or minutes, and that agents can then run several instruments in parallel — reasoning through an experiment, adjusting parameters mid-run, and in some cases recovering from a hardware error without a human.

The named early work is specific enough to check: an automated protein assay at Genentech, protein-design screening at the University of Washington Baker and Pinglay labs, dose-response curves at Carnegie Mellon, microscopy at HHMI Janelia, laser stabilisation on a quantum computer at QuEra Computing, and environmental qPCR at Tetsuwan Scientific. Access is by application at modelhardwarestandard.com. Anthropic says it is developing safety evaluations with partners ahead of open-sourcing the standard.

Two properties matter more than any of the above, and both are easy to read past.

It works with any device that has a programmable interface, and it is model-agnostic — any agent harness can reach it using standard protocols including the Model Context Protocol.

Anthropic, in other words, built the socket and declined to require its own plug.

The tell

That is not how a company enters a market. It is how a company defines one.

Compare the two obvious competitors. Nvidia’s Isaac GR00T N1.7 is a 3-billion-parameter open vision-language-action model with commercial licensing, pretrained on more than 20,000 hours of human egocentric video — and Nvidia reports a dexterity scaling law, where going from 1,000 to 20,000 hours of that data more than doubles average task completion. Google’s Gemini Robotics is a VLA model adapted across ALOHA, Franka and Apptronik hardware. Both companies are selling the brain.

Anthropic shipped no brain. It shipped the wiring, and its own people describe it in exactly those terms — one member of technical staff called the protocol’s role “the USB for AI to software connection,” while the head of partnerships framed the goal as “we want to avoid vendor lock-in for scientists.”

The precedent is close enough to be a template. MCP began the same way: an Anthropic-authored spec, given away, deliberately model-agnostic, marketed with a USB analogy. It went from roughly 100,000 SDK downloads in its launch month to about 97 million a month by March 2026. In December 2025 Anthropic donated it to the Linux Foundation’s Agentic AI Foundation, and OpenAI, Google, Microsoft, AWS, Cloudflare and GitHub built on it. Anthropic captured no licensing revenue from any of that. What it captured was the shape of the interface — and a position where every agent product on the market, including its competitors’, speaks a dialect it wrote. The Stainless acquisition was the same instinct applied to SDKs.

MHS repeats every structural choice: spec not product, model-agnostic not Claude-only, open source promised not licensed, and launched with hardware vendors rather than customers.

The part the coverage skipped

Here is what almost none of the launch write-ups mentioned. Lab instrument integration is not an unsolved problem for lack of a standard. It has two.

SiLA 2 is a mature open standard for laboratory device connectivity, built on gRPC with a Feature Definition Language that already supplies machine-readable, typed descriptions of instrument commands and observable properties. OPC UA, with its LADS companion spec, does the analogous job on the industrial side. Both are open. Both are years old. Both were designed to end exactly the bespoke-integration misery Anthropic’s announcement describes. An academic effort in the same direction — a June 2026 preprint proposing LAP, an agent-to-instrument protocol — explicitly positions itself as complementing SiLA 2 rather than replacing it.

And laboratories still are not plug-and-play. The documented reason is not protocol quality. It is that a lack of interface standardisation among hardware vendors leaves most devices exposed only through proprietary interfaces, so every integration needs dedicated development. OPC UA’s own specification complexity produced incomplete and heterogeneous implementations, which made writing generic software against it hard in practice even where it was nominally supported.

That is a vendor-incentive failure, not an engineering one. An instrument maker that ships a great open interface makes its hardware easier to replace.

MHS does bring one genuine design difference: SiLA 2 describes a device to a programmer writing a deterministic workflow, while MHS describes it to a model that has to reason about it — hence natural-language capability and constraint tags alongside the typed layer, and hence riding on MCP, which agent harnesses already speak. That is real. It is not sufficient. Nothing in a better description changes what a vendor gains by staying closed.

Which is why the most consequential line in the announcement is a list of company names. AWS, Tecan, QIAGEN and Universal Robots are adding MHS support. Tecan and QIAGEN are precisely the incumbents whose proprietary interfaces constitute the problem. If that list grows over the next two quarters, MHS succeeded where SiLA 2 stalled. If it does not, the spec’s quality will be irrelevant.

Why the MCP curve probably doesn’t repeat

MCP’s growth had a specific engine: software integration is nearly free to attempt. Write a server in an afternoon, run it, and if it is wrong it returns bad data. That cheap failure mode is what allowed thousands of teams to try it simultaneously.

Physical integration has no such property. A wrong driver breaks a six-figure instrument, destroys an irreplaceable sample, or hurts somebody. Anthropic’s own materials concede the boundary — Claude requires expert oversight for physical and chemical constraints, and safety limits are enforced at the device level rather than left to the model.

You can see that asymmetry in how the two were launched. MCP arrived as a public spec with open SDKs. MHS arrived as an application-gated preview whose stated purpose is developing safety evaluations before open-sourcing. Expect a slower curve, gated on driver support and safety validation rather than on developer enthusiasm.

The same week, the same move

Two days earlier, in Nvidia’s reported acquisition of Hugging Face, the neutral open-weight distribution layer was bought outright. Hugging Face then launched Microduck, a $399 open-source robot, on the same day as the MHS preview — and notably not using MHS.

The pattern connecting them is one this desk has tracked all month: the fight has moved from models to the layers between them. OpenAI is consolidating the harness layer. Orchestration tooling is going deliberately model-agnostic. Nvidia bought a distribution layer. Anthropic is trying to author a hardware layer.

The buyer-relevant lesson from that same month is a caution, and it comes from the Assistants API sunset: vendor-owned integration layers have a half-life, and when they are retired the migration cost lands on whoever built against them. An open, foundation-governed standard is materially safer than a proprietary one — and MCP’s donation to the Linux Foundation is evidence Anthropic will follow through. But MHS is not there yet. Today it is a gated preview under a single company’s control, with open-sourcing promised rather than done. That gap is the whole risk, and it closes only when the spec actually ships to a neutral home.

What this changes for what you buy

If you do not operate laboratory or manufacturing instruments: nothing. No price moved, no model shipped, no product became purchasable. Your choice among Claude, Claude Code or anything in our AI agents guide is untouched by this announcement.

If you do operate instruments, there is one action, and it is free: put MHS support on the question list for your next hardware procurement or renewal. Tecan, QIAGEN, Universal Robots and AWS have committed to it. What your vendors say when asked — and whether they can name a shipping date — is the most reliable early signal available on whether this standard becomes real.

If you are planning agent infrastructure more broadly, the useful read is directional. Anthropic has now chosen own the interface, not the application three times running. That is a consistent enough pattern to plan around: expect the integration surfaces your agents use to keep converging on MCP-shaped, foundation-governed specs, and treat any vendor-proprietary alternative as carrying migration risk by default.

The read

The interesting thing about MHS is not that Anthropic touched hardware. It is that the company with the strongest agent models on the market responded to physical AI by shipping no model at all.

Nvidia and Google are competing to sell the best robot brain. Anthropic looked at that fight and went one layer down, to the socket both of those brains eventually have to plug into — and gave it away, exactly as it did with MCP, on the theory that owning the shape of the connection is worth more than winning any particular model generation.

That theory has been right once already, at very large scale. Whether it is right twice depends on something Anthropic does not control and cannot engineer around: whether a handful of instrument manufacturers, whose commercial interest has pointed the other way for twenty years, decide this time to open up.

The spec is the easy part. It always was.

Questions or corrections? Email Pick Right. Want the full list? See all news.