Anthropic opened a research preview of the Model Hardware Standard on August 27, 2026, and the shorthand that formed around it within a day was "MCP for physical devices." That shorthand is wrong in a way that matters if you are going to build against it. This piece covers what the specification actually is, where MCP genuinely fits, what a device file declares, who is testing it, the limitations Anthropic published about its own model, and the European regulation that lands roughly seventeen months from now and may reclassify the whole thing.
Anthropic is shipping a driver, not a protocol
The announcement opens with the broad framing, "a shared specification for AI agents to safely operate physical devices," and then gets specific in a way most coverage skipped. It calls the thing "a standardized driver: software that translates between a computer’s operating system and a hardware device."
Driver is a precise word and Anthropic chose it deliberately. A protocol defines how two parties talk. A driver defines what one particular piece of equipment is, what it will accept, and what it will refuse. That difference is not academic. It determines who writes the file, who owns it, and who is answerable when it is wrong.
The problem being solved is described in plain operational terms. "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. MHS reduces this integration work to hours or minutes."
Jonah Cool, Anthropic’s head of partnerships and deployment of science, gave Fortune the motivation directly: "We want to avoid vendor lock-in for scientists," noting that scientific equipment "suffers from proprietary solutions that are very brittle."
Where the model hardware standard actually sits relative to MCP
Here is the sentence that dissolves the shorthand. On interfaces, the announcement says, "There are three such mechanisms: MCP, the command line interface, and code files (APIs). These work together to enable orchestration across multiple devices via a single line of code."
MCP is one of three doors. Not the building.
The announcement reinforces the point elsewhere: "MHS is also model-agnostic, and any agent harness can access it using standard protocols, such as the Model Context Protocol." Standard protocols, plural, with MCP given as an example rather than as the mechanism.
| Layer | What it does | Who authors it |
|---|---|---|
| The device itself | Exposes a programming interface, or does not | Hardware vendor |
| MHS driver file | Declares readable values, writable values and enforced safety limits | Vendor or integrator |
| Access mechanism | MCP server, command line, or a code file calling an API | Whoever is wiring up the agent |
| Agent harness | Plans and issues the calls | Model-agnostic by design |
Two consequences follow. If you have already invested in MCP servers, you do not get MHS for free, because the driver layer underneath is the part that did not previously exist. And if you have no MCP at all, you are not locked out, because a shell command or a direct API call reaches the same driver.
The permission scoping questions we raised in our piece on Antigravity and MCP apply with more force once the thing on the other end moves.
What a device file declares
The primitives are deliberately small. The spec uses "a simple set of primitives, commands like ‘read’ (for example, ‘get temperature’) or ‘write’ (for example, ‘set temperature’), that any hardware device can understand and act on."
Discovery is handled at the same layer. Devices become "discoverable in a standard format, so that devices and agents can find each other and communicate across networks without needing a bespoke ‘translator’ program in between."
The authoring step is the interesting one. The driver "contains tags that let the user write this information directly in natural language," and from those tags the toolchain produces "a reference file with information about a device’s general characteristics, such as what it can measure, what can be adjusted, and what safety limits will be enforced."
Read that last clause again, because the rest of this story hangs on it. The file does not merely describe a device. It declares the limits that will be enforced on an agent driving that device. Anthropic member of technical staff Alek Kemeny called MCP "kind of like the USB for AI to software connection." The model hardware standard is the layer below that, where the equipment states what it will and will not permit.
Who is actually in the preview
Coverage has collapsed two different lists into one, making the program look broader than it is. The announcement separates them.
| Group | Named participants |
|---|---|
| Research labs and scientific users | Genentech, the University of Washington Baker and Pinglay labs, Carnegie Mellon University, HHMI Janelia Research Campus, QuEra Computing, Tetsuwan Scientific |
| Hardware and platform partners | Amazon Web Services with Strands Robots, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots |
| Also named | Hugging Face, Raspberry Pi |
Fortune reports that only "a handful" of labs and hardware manufacturers received early access during this phase. Access is by application through modelhardwarestandard.com, which states that "MHS is starting as a limited research preview" and that access "is by application during the preview."
There is a hardware precondition that will exclude a lot of existing equipment. Fortune notes that not all installed devices can connect out of the box, because not all of them have a programming interface at all. Anthropic says it is working with manufacturers both to build new products with the necessary interfaces and to retrofit existing ones. That is a multi-year hardware refresh cycle sitting underneath a software announcement.
On open sourcing, the language is intent without a date. Anthropic says it is sharing an early version with partners "ahead of making the standard open source," and that when it does, it "will release findings from the research preview as part of our guidance for deploying the standard safely." No timeline is given.
Claude paused overnight waiting for permission
The most useful section of the announcement is the one describing what did not work, and it is worth noting that Anthropic published these results itself rather than leaving them for a reviewer to find.
The structural limitation is stated plainly: "As a large language model, Claude learns about the physical world through text and images, meaning its spatial and physical reasoning have limitations that still require expert oversight." The field results match. "Although they excel at general-purpose reasoning, they still struggle with physical, chemical, and biological constraints, particularly when troubleshooting errors that call for real-world physical intuition."
Then there is the detail that anyone who has run an overnight experiment will recognize immediately. A partner reported that "Claude often stopped to wait for human confirmation before performing an action it deemed even slightly risky, meaning experiments would sometimes pause overnight while Claude waited for approval."
At Genentech, working on bubble formation during liquid handling, the team reported that "we had to guide it towards parameters that handled the liquid more gently."
None of that is a failure of the specification. It is an accurate picture of where the capability sits, and it is the opposite of the autonomous laboratory framing much of the coverage ran. The behavior is closer to a careful new technician who escalates constantly, which is the right default for equipment that can spill, heat, or swing. Our practitioner’s guide to AI agents covers why that escalation frequency is a design choice rather than a defect.
In Europe, that file may be a regulated safety component
This is the part with a deadline attached, and almost nobody covered it.
Regulation (EU) 2023/1230 replaces the Machinery Directive. According to EU-OSHA, "The Regulation applies from 20 January 2027. However, some rules already apply on earlier dates such as the requirements for notified bodies (20 January 2024)." The same source describes the scope expansion: the regulation "intends to better cover new technologies such as autonomous mobile machinery (robots), internet of things with connected equipment, or artificial intelligence (AI), where specific modules of AI using learning techniques ensure safety functions."
The Next Web drew the line to MHS first. Its argument is that a vendor can specify in an MHS file how an AI may move a heavy arm safely, "limiting the speeds and angles it is allowed to use," and that "a specification that constrains a robot arm’s speed and angles is performing a safety function." The conclusion follows: "In Europe, whoever writes that file may find they have written a regulated safety component."
That reading is an interpretation, not a settled position, and no regulator or notified body has tested it. But it is not a stretch, and the sequencing is awkward. The preview opened in August 2026, open sourcing has no date, and the regulation applies in January 2027. A standard still private when the obligation arrives is hard for a European manufacturer to build a conformity assessment around.
What to check now if you build or buy this equipment
Access is gated, so nothing here demands action this quarter for most teams. The questions worth settling early are ownership questions rather than technical ones.
| Question | Why it matters |
|---|---|
| Who authors the device file, you or the vendor? | Determines who owns the declared safety limits |
| Does the equipment have a programming interface at all? | Devices without one cannot participate without a retrofit |
| Which of the three access mechanisms will you standardize on? | MCP, command line and API files have different audit surfaces |
| What is your policy when the agent pauses for approval? | Overnight stalls were reported by preview partners |
| Do you sell into the EU? | January 20, 2027 is a fixed date and the standard is not yet open |
Settle the safety limits question first. A limit declared in a vendor file behaves very differently from one enforced in device firmware, and the specification does not remove the need for the second. The lessons from agentic browser security transfer directly, because there too a control that looked like a hard boundary turned out to be a declaration something else had to honor. What changes here is that the failure mode stops being data loss.
For context on where this fits in the broader hardware picture, our survey of physical AI agent devices maps the consumer and industrial sides of the same trend.
Frequently Asked Questions
Is the Model Hardware Standard the same thing as MCP?
No. Anthropic describes MHS as a standardized driver and lists MCP as one of three ways to reach it, alongside a command line interface and code files exposed as APIs. It sits beneath the access mechanism rather than replacing MCP.
Can I use it today?
Only by application. It is a limited research preview open to a first group of research labs and advanced manufacturers, and Fortune reports only a handful have early access so far.
Will it be open source?
Anthropic states the intention clearly and gives no date. The announcement says it is sharing an early version with partners ahead of making the standard open source, and that findings from the research preview will be released as deployment guidance when it does.
Does it only work with Claude?
No. The announcement states MHS is model-agnostic and reachable by any agent harness using standard protocols. All published field results come from Claude, so that is a design property rather than something demonstrated across vendors yet.
What does a device file actually contain?
It is generated from natural language tags and produces a reference file covering what a device can measure, what can be adjusted, and what safety limits will be enforced. The command surface is a small set of read and write primitives.
Will my existing lab equipment work with it?
Not necessarily. Fortune reports that not all existing equipment can connect out of the box, because not all of it has a programming interface. Anthropic is working with manufacturers on new products and retrofits, which puts some of this on a hardware timeline.
What is the European regulatory issue?
Regulation (EU) 2023/1230 applies from January 20, 2027 and covers AI modules performing safety functions. The Next Web argues that a file constraining a robot arm’s permitted speeds and angles is one, making its author responsible for a regulated safety component. That reading is untested by a regulator.
How autonomous is this in practice?
Less than the coverage suggests. Anthropic reports that Claude’s spatial and physical reasoning still require expert oversight, that it struggles with physical and chemical constraints when troubleshooting, and that one partner saw experiments pause overnight awaiting human approval.
