On 27 August, Anthropic previewed the Model Hardware Standard — "a shared specification for AI agents to safely operate physical devices." Not a product. A common language.

Right now a machine can usually only be run by its own vendor's software. That's why getting two machines from two suppliers to work together takes a specialist and, in Anthropic's estimate, weeks or months. MHS gives every device the same kind of driver, lets it announce itself on the network so agents and other machines can find it, and carries the machine's capabilities and its safety limits in plain English rather than code. Anthropic's claim is that this drops integration to hours or minutes.

It's a research preview — a limited set of partners, and not open-sourced yet.

Hashi's take: new AI frontier models are released every month, it seems. But when MCP — and now MHS — comes along, that makes a meaningful shift in the value creation of AI.

If your operations lead still treats integration as a line item rather than a strategy, forward this their way.

Anthropic's announcement opens with the problem rather than the technology: "It typically takes a lab or manufacturing facility weeks, if not months, to set up and integrate their hardware. Most devices don't communicate with each other, instead requiring specialists to build bespoke integrations."

None of that is an AI problem. It has been true for decades. Machines come from different vendors, each vendor ships its own control software, and connecting any two of them is a project with a budget and a consultant attached.

MHS is their answer. One agreed way for a machine to describe what it is and what it can do, so software can operate it without anyone writing custom code for that particular model. Anthropic says this "reduces this integration work to hours or minutes."

It's a research preview. A limited set of partners have access by application, and the standard isn't open-sourced yet.

A standard driver. Same idea as a printer driver — software that sits between the computer and the device and translates. Except instead of a different driver disc for every model, MHS defines one shape for all of them, built on simple commands like read and write.

Discovery. Devices publish themselves on the network in a standard format, so an agent and a machine can find each other without a translator program written specially for that pairing. Your equipment stops being invisible.

Plain-English tags. This is the part I really like. The driver carries a description of what the machine can do and what its safety limits are, and that description is written in natural language rather than code. The centrifuge says what it is, what it can do, and what you must never ask it to do — in English.

Three ways to control it. MCP, the command line, or code files. Whichever your team already works in. That's the technical layer, and it belongs to your IT people rather than to you.

This isn't about making the equipment smarter. It's about putting a translator in the middle — one agent, one language, and every machine it needs to reach.

In November 2024 they published the Model Context Protocol, and the problem it described was the same one. Models were "constrained by their isolation from data — trapped behind information silos and legacy systems," and "every new data source requires its own custom implementation, making truly connected systems difficult to scale."

Swap "data source" for "machine" and you have the MHS announcement.

Based on MCP's success to date, MHS could see a similar trajectory. By Anthropic's own count, more than 10,000 active public MCP servers and 97 million monthly SDK downloads. Adopted by ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code — competitors, all of them, using the same plumbing. Then in December 2025 Anthropic packaged the whole thing up and handed it to a fund at the Linux Foundation, co-founded with Block and OpenAI and backed by Google, Microsoft and AWS. You give a standard away when you want everyone building on it.

Every time you or someone on your team has connected an outside app to Claude or ChatGPT, that connection almost certainly runs on MCP. You probably never learned its name or thought about it twice. And it just stays in the shadows.

MHS will likely end up in the same place. Then one day the lab equipment and the machinery on your floor are talking to each other, and AI is what's doing the translating.

Anthropic published six partner results with the preview. Two of them stand out to me.

QuEra Computing runs quantum computers, which need lasers held at exact frequencies. Stabilising them used to succeed 58% of the time. With Claude operating the equipment through MHS, that rose to 99.3%. When it did drift, recovery went from 150 seconds to six.

Genentech automated one of its lab tests — the kind that runs across several different instruments and normally needs someone shepherding the sample between them. Claude ran the sequence and tuned the liquid-handling settings itself as it went, rather than following numbers fixed in advance.

That last part is the unlock. Most automation executes a set sequence with no view on whether the output was any good. This adjusted as it went.

You probably don't run a quantum computer or a research lab. What carries over isn't the science, it's the shape: equipment from different vendors, each with its own software, and a skilled person in the middle getting them to cooperate. That person is the bottleneck and the single point of failure. Every plant, clinic, warehouse and test lab has one.

Getting two machines to work together has always taken lots of time and lots of money. So most testing runs with people in the gaps — someone reads one instrument's output, sets it up on the next, carries the sample across. It's slow, and every handoff is a place where something can go wrong.

Take that work from weeks to an afternoon and the gaps close. QuEra's numbers are the clearest version of what that buys: not just faster, but more reliable.

The labs running these instruments are the ones working on drugs, materials, batteries, diagnostics. If an experiment that took a week takes a day, and fails less often for reasons that have nothing to do with the science, then discovery gets quicker and it gets cheaper. Not at one lab. Across all of them at once.

The prize isn't a faster lab. It's faster medicine, cheaper materials, and inventions that show up years before they otherwise would.

Anthropic published the limitations alongside the preview. The gist:

None of this bothers me much. MCP was a lot less capable at launch than it is now, and so was the first model any of these labs shipped. The improvement curve is steep enough that I'd bet on most of these limitations looking very different in three to six months.

Last November, I wrote about Bezos taking over Project Prometheus. My point of view was that internet text has been mined out, and the next real advantage comes from data your own operations produce. That point hasn't changed for me, and MHS is what unlocks it. Back then I put it this way: "unlike digital AI where everyone trains on the same internet data, physical AI creates natural moats. The company with seven years of warehouse robot data has an advantage you can't replicate."

Then in July, writing about single-serving AI, I said the small devices turning up everywhere wouldn't stay isolated for long — "it's only a matter of time before they connect." I assumed that connection would arrive from the consumer side, since that's where the devices were multiplying. It's coming from labs and factories instead.

The missing link was that you can't generate operational data from equipment that can't be read, and you can't act on it with equipment that can't be told anything. To build a moat you need the right tools, and MHS is that tool. Nothing in the specification is specific to lab equipment, so I'd expect the consumer side to follow before long.

Keep reading!

- Hashi