Summary: Anthropic announced a research preview of the Model Hardware Standard on August 27, 2026 to let AI agents operate lab and factory equipment safely. Standardized drivers, discoverability, natural-language device tags, MCP, and code files are intended to reduce integration time from weeks toward minutes. Early partner projects and vendor support are encouraging, but physical reasoning gaps and hardware without programmable interfaces remain limits.
The Shift: A Standard for AI-Controlled Hardware
Wiring lab or factory equipment together usually eats weeks, if not months. Every instrument comes with its own programming interface, so the integration code has to be written to order, and devices have no shared method for handing data over to an AI agent. On August 27, 2026, Anthropic announced a research preview of the Model Hardware Standard (MHS) — a shared specification that lets AI agents safely run physical devices such as microscopes, liquid handlers, and robotic arms. The aim is to shrink integration time from weeks down to hours or minutes.
MHS started out as a joint effort between Anthropic and HHMI Janelia Research Campus. Today it is open to a first cohort of scientific research labs and advanced manufacturers. Any device with a programmable interface can use the standard, which is model-agnostic and reachable through standard protocols such as the Model Context Protocol (MCP).
Inside MHS: Standardized Drivers and Natural Language Tags
At the center of MHS sits a standardized driver: software that translates between a computer’s operating system and a hardware device. That driver relies on a small set of primitives — commands such as “read” (e.g., “get temperature”) or “write” (e.g., “set temperature”) — which any hardware device can understand. The same driver also makes each device discoverable in a standard format, so devices and agents can locate one another across networks without any bespoke translator.
Just as important, the driver helps an AI agent make sense of a device it has never encountered before. Built into it are tags that let users describe machine characteristics in natural language — the weight of a robot arm, for instance, which matters for safe manipulation. Users can write these themselves, or they can chat with an agent that interviews them about their hardware. The driver then turns those tags into a reference file automatically, covering general characteristics: what the device can measure, what can be adjusted, and which safety limits are enforced. With that file in hand, the agent has everything required to operate the device.
With devices connected, three control mechanisms come into play: MCP, command line interface, and code files (APIs). They can be combined, letting a single line of code orchestrate several devices at once. When a task runs long, or when operations move faster than online reasoning can keep up with, the agent can chain driver commands into code files, so the devices carry out the work on their own without step-by-step reasoning.
Early Partner Results: Protein Assays to Quantum Lasers
Anthropic handed MHS to partners working in biotech, robotics, quantum computing, and other areas. A few of the early projects:
- Genentech stood up MHS as a proof-of-concept to automate the BCA protein assay, coordinating a liquid handler, a robotic arm, and a plate reader. Claude orchestrated the protocol and tuned the liquid handling parameters on its own. It settled on an optimal flow rate of ~140 µL/s (0.016 RMSE) for water and ~10 µL/s (0.181 RMSE) for viscous BSA — figures automation experts judged reasonable. Claude also pulled itself out of errors such as tip pickup failures.
- University of Washington Baker and Pinglay labs PhD student Zihao Song used MHS to build three things: a remote monitoring dashboard, an AI agent-supervised qPCR that tracks amplification curves and stops at the right moment, and a collision-free integration between a robotic arm and a liquid handler.
- Carnegie Mellon University researchers turned to MHS for serial dilution dose-response experiments and ran them roughly three times faster than before, coordinating a liquid handler, plate reader, robotic arm, and monitoring cameras spread across three incompatible computers.
- HHMI Janelia scientist Virginie Ruetten used MHS to bring a microscope setup under one roof and coordinate it; that setup had previously demanded seven different vendor programs.
- QuEra Computing put an AI agent in charge of the laser systems inside a quantum computer. The controller the agent developed restored laser lock on its own in 99.3% of cases, with no human intervention.
- Tetsuwan Scientific folded MHS into ResearchOS, its automated biology platform, for qPCR workflows that support citizen science pollution analysis in San Pedro Creek, California.
Hardware vendors are signing on too. AWS will deliver MHS support through Strands Robots. Automata is adding MHS to its LINQ platform. Danaher is exploring smart instruments alongside Anthropic. Doosan Robotics is testing robotic arm coordination. MBF Bioscience is writing a driver for ScanImage. QIAGEN is running a proof-of-concept on QIAsymphony Connect. Tecan and Universal Robots both plan to support it.
Limits and Trade-offs: The Physical Reasoning Gap
The limitations of MHS are plain to see. As a large language model, Claude takes in the physical world only through text and images, which leaves its spatial and physical reasoning constrained and dependent on expert supervision. Genentech researchers, for instance, had to steer Claude toward recognizing that bubbles in the samples were a physical problem rather than a software error. Claude’s first instinct was to retry with different parameters, which merely stirred the fluid further. Told about the physics, it adjusted and held onto that context.
Hardware without a programmable interface is another gap: MHS does not support it. Anthropic is teaming up with makers of those devices to build drivers in. Throughout the research preview, Anthropic will work with partners on safety evaluations and on a physical safety roadmap aimed at guarding against misuse. Plans call for open-sourcing MHS in the future, at which point findings from the research preview will be released as guidance for safe deployment.
What Builders Can Take Away
For product builders and anyone learning AI tooling, MHS lays out a concrete route for connecting AI to physical devices. The central insight: standardizing drivers and leaning on natural language tags can cut integration time sharply. Rather than writing custom code for every device, you describe device characteristics in plain language and let the driver produce a reference file for the agent.
A few things to weigh when you build with MHS:
- Start with a programmable device: any device with a programmable interface works with MHS. If your hardware lacks one, you’ll have to work with manufacturers or add a driver.
- Use natural language tags: enrich your device drivers with tags describing physical characteristics. They help the agent grasp what safe operation looks like.
- Reach for code files on long tasks: when operations demand speed or stretch over long durations, chain driver commands into code files so devices can run autonomously.
- Plan around the physical reasoning gaps: don’t count on the agent to handle every physical nuance. Build in expert oversight and feedback loops so errors get corrected.
Applications for the research preview are open at modelhardwarestandard.com. If you build tools that interact with lab or factory equipment, this is a promising standard to keep an eye on.
Bottom Line
MHS marks an early yet promising move toward making AI agents practical for automation in the physical world. It shortens integration time, makes parallel operation of multiple devices possible, and lets agents learn and adapt. It is no silver bullet, though: physical reasoning remains a challenge, and expert supervision is still necessary. As the standard evolves and goes open source, it could unlock new levels of autonomy in labs and factories, but builders should come in with realistic expectations about where its limits sit today.