For years, AI agents have lived entirely on screens — writing code, answering questions, and navigating the digital world with increasing sophistication. That era may be drawing to a close. On August 27, 2026, Anthropic made an announcement that could fundamentally reshape how AI interacts with physical infrastructure: the Model Hardware Standard (MHS), a new open specification designed to let AI agents safely discover and operate real-world machinery. From laboratory microscopes to robotic arms to quantum computers, the physical world is about to get a universal language for talking to AI — and the implications are profound.
What Is the Model Hardware Standard?
Anthropic announced the Model Hardware Standard, or MHS, a new interface that will make it simpler for AI agents to operate and communicate with physical machinery.
Think of it as a universal translator between AI models and the physical devices they need to control.
The company describes it as working like a USB-C cable, standardising the way information passes between a model and a device.
Just as USB-C eliminated the chaos of proprietary charging connectors, MHS aims to eliminate the fragmented, bespoke software integrations that currently make connecting AI to hardware so painful and time-consuming.
At the core of MHS is a standardised driver that acts as software between a computer's operating system and a physical device. The MHS driver uses a set of basic commands, or primitives, including "read" and "write." A read command could be used to obtain a device's temperature, while a write command could set the temperature. The system also makes devices discoverable in a standard format, allowing hardware and AI agents to find and communicate with one another across networks without requiring a bespoke translator for each device.
The idea builds on a project Anthropic already shipped for software — the Model Context Protocol (MCP), which gives AI models a standard way to access external data sources and tools. MHS extends that approach to machines.
The Problem MHS Is Solving
To understand why this matters, you need to appreciate just how broken the current state of hardware integration really is.
Today's industrial and scientific equipment often relies on proprietary interfaces and specialised integration work. That can make deploying AI systems across different machines expensive and time-consuming, because developers have to build separate connections for each piece of hardware.
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.
The genesis of MHS illustrates the problem perfectly.
Alek Kemeny, of Anthropic's Beneficial Deployments team, began the MHS project with Arco Bast, a postdoctoral scientist at HHMI Janelia. Bast was running brain-imaging experiments on a rig that combined lasers and motorised focusers, and the setup also used cameras from different vendors with no shared interface.
The result was a patchwork of incompatible systems that slowed scientific progress to a crawl.
If MHS can reduce integration work from months to hours or minutes, as Anthropic argues, it could change who gets to automate sophisticated experimental setups. Today, that ability often depends on unusually strong in-house engineering talent. A standard interface could make automation more portable across labs and institutions, lowering friction for teams that want to connect instruments, orchestrate runs, and capture structured data more consistently.
Real-World Results: What the Early Tests Show
This isn't vaporware. Anthropic has been testing MHS with a select group of partners, and the results are striking.
In early testing, Claude used MHS to improve laser stabilisation on QuEra's quantum computers from a 58% to 99.3% success rate, cut a multi-day imaging workflow at HHMI Janelia down to a single day, integrate six lab instruments at the University of Washington in under a week, and run an automated protein assay at Genentech that recovered from real equipment failures — tip pickup errors and fluid detection faults — without a human stepping in.
Researchers at Carnegie Mellon University used MHS to run serial dilution dose-response experiments about three times faster than before, with an AI agent orchestrating a liquid handler, a plate reader, a robotic arm, and monitoring cameras spread across three computers with fundamentally incompatible interfaces.
HHMI Janelia unified seven vendor programs into a single interface, with new hardware integration time dropping from multiple days to minutes. The University of Washington connected six lab instruments in under a week — replacing months of traditional automation work — and eliminated manual plate swapping every 90 minutes during qPCR runs.
These are institutional research partners running real experiments, not curated demos. The numbers speak for themselves.
Who Is Already on Board?
Organisations testing or developing support for MHS include Amazon Web Services, Automata, Danaher, and Doosan Robotics. AWS plans to support MHS through Strands Robots, while Hugging Face is adding support to LeRobot and Raspberry Pi is enabling integrations across several products.
One critical design decision sets MHS apart from a typical vendor play:
a standard controlled exclusively by one AI company would have limited appeal to manufacturers and enterprises that want flexibility over which models they deploy. By making MHS model-agnostic and eventually open source, Anthropic is attempting to position it as infrastructure rather than simply another Claude feature.
Reporting on the launch notes that the standard is open to any model, not just Claude — which is the detail that will determine whether the rest of the industry adopts it or builds a competing one.
This is Anthropic playing a long game, betting that owning the standard matters more than owning the only model that can use it.
The Safety Question: Proceeding With Caution
Giving an AI agent control over physical machinery introduces an entirely different category of risk compared to letting it write an email or generate a report.
Giving an AI agent control over physical equipment raises a different class of risks from letting it write code or manipulate information. A mistake can damage machinery, ruin an experiment, or potentially create safety hazards. A compromised agent, account, or driver could also expose connected equipment to unauthorised commands.
The early tests also surfaced important limitations in the AI's reasoning.
Genentech researchers had to guide Claude to see foaming in protein samples as a physical failure and not a software bug — one that needed a physical correction instead of a code fix. Claude's spatial and physical reasoning carries limits, Anthropic notes, because the model learns about the physical world through text and images, not through direct sensing.
At QuEra, failures of the physical hardware itself could not be resolved by AI, and whenever an operation was judged even slightly risky, the system waited for human approval — sometimes halting an experiment overnight. A design that cautiously stops avoids accidents, but it trades off against the uptime achievable with unattended operation.
Anthropic is taking these challenges seriously.
The company has committed to using the research preview to build additional safety evaluations with its launch partners and strengthen protections for the use of AI in the physical world, and is developing a physical safety roadmap to further bolster its safeguards policy and enforcement coverage against the risk of misuse.
Before adopting MHS beyond controlled tests, organisations will need to determine whether it can enforce least-privilege access, record agent activity, and fail safely when an agent or connected component behaves unexpectedly. Security teams should treat AI-to-hardware interfaces as critical control infrastructure rather than ordinary software integrations.
Practical Tips: How to Prepare for the Age of Physical AI Agents
Whether you're running a research lab, managing a manufacturing floor, or building the next generation of agentic software, here's how to get ahead of this shift:
- Join the waitlist now.
Anthropic has opened the MHS research preview to a first group of scientific research labs and advanced manufacturers.
Applications are open, and early participants will shape the standard's future direction.
- Audit your hardware for programmable interfaces.
Anthropic acknowledges that MHS does not yet work with hardware lacking a programmable interface, and the company is working with manufacturers of such equipment to build in drivers.
Start by cataloguing which of your devices already have APIs or programmable control layers.
- Start with low-stakes, reversible workflows.
Scope pilots to reversible or low-consequence workflows first, and keep human oversight in the loop for anything where a mistake is not undoable.
This is the safest way to build confidence in AI-driven hardware control without catastrophic downside.
- Avoid vendor-specific lock-in.
The durable payoff of MHS is a standardised driver ecosystem that compounds in value as adoption grows, exactly as MCP did in software. Teams building agent infrastructure for research or manufacturing should evaluate MHS compatibility now, before vendor-specific integrations lock in technical debt.
-
Treat AI-hardware interfaces as security-critical infrastructure. Implement authentication controls, audit logging, and network isolation for any system where an AI agent can issue commands to physical equipment.
-
Follow the open-source release.
Anthropic plans to publish research findings and safety guidance when the standard is open-sourced.
That documentation will be essential reading for any team planning a serious deployment.
The Bigger Picture
The announcement signals that Anthropic is wading deeper into hardware, where rivals including OpenAI and Amazon have spent billions of dollars designing AI-native devices and manufacturing tools.
The race to control how AI interacts with the physical world is accelerating, and MHS represents Anthropic's most ambitious bid yet to define the rules of engagement.
The project frames standardisation as a prerequisite for AI agents to act in the physical world
— and that framing is exactly right. The same way MCP unlocked an explosion of software integrations for AI agents, MHS could do the same for factories, labs, hospitals, and industrial facilities worldwide.
Conclusion: The Physical World Just Got a New Protocol
Anthropic's Model Hardware Standard isn't just a developer tool — it's an infrastructure bet on where AI is heading. The early results from Genentech, Carnegie Mellon, HHMI Janelia, and QuEra demonstrate that AI agents can already deliver measurable, real-world value when given a standardised way to talk to physical machinery. The remaining challenges around physical reasoning and safety are real, but they're solvable — and Anthropic is actively working to solve them in public, with partners.
If you're a researcher, engineer, operations leader, or technology decision-maker, now is the time to pay attention. The organisations that learn how to work alongside AI-controlled physical infrastructure in the next 12–18 months will have a significant head start on everyone else. Visit anthropic.com to explore the MHS research preview, join the waitlist, and start evaluating where physical AI agents could transform your workflows today.
