On August 27, 2026, Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared protocol that lets AI agents operate physical machines. The announcement covers lab instruments, robotic arms, and factory equipment, and it signals Anthropic’s most direct push yet from software into the physical world.
The release has immediate relevance for enterprise teams building on AI in any industry that touches real-world hardware, from pharmaceutical research to industrial manufacturing to quantum computing.
What MHS Is and Why the Analogy Is Accurate
Anthropic describes MHS as the USB-C of physical hardware for AI agents. The comparison is apt. USB-C standardized how data and power move between devices so manufacturers stopped designing proprietary connectors. MHS does the same thing for AI-to-machine communication: it defines a common driver layer so any AI agent can discover and operate any MHS-compatible device without custom integration work.
Before MHS, connecting an AI agent to a lab instrument required weeks of bespoke engineering per device. Most instruments do not communicate with each other natively, so every integration was a one-off. Anthropic says MHS compresses that timeline from weeks or months to hours or minutes.
The mechanism: each physical device gets an MHS driver that exposes a small set of standard primitives, read and write operations similar to the commands used in a Bash shell. The driver generates a reference file describing device characteristics and capabilities. An AI agent can then discover the device, reason about what it does, issue commands, monitor outputs, and adjust parameters in real time.
Critically, MHS is model-agnostic. Agents built on Claude, GPT, Gemini, or open-weight models can all use the same driver. The standard routes through three control paths: the Model Context Protocol (MCP), the command line interface, and a direct API. Enterprise teams that have already adopted MCP for software integrations will find the pattern familiar.
How MHS Relates to MCP
Anthropic introduced MCP in 2024 as a standard for connecting AI models to software data sources. It has since been adopted far beyond Anthropic’s own stack, becoming a de facto protocol for agentic tool integrations across the industry. MHS is the physical-world counterpart.
The design is intentionally parallel. MHS lets agents interact with hardware through MCP, which means an agent already using MCP for software tools can extend its reach to connected physical equipment through the same channel. The plumbing is the same; the endpoint is now a machine rather than a database or an API.
This approach mirrors the open-source playbook Anthropic used with MCP: release a standard, build ecosystem adoption, and let the value compound as more devices and model vendors implement it. If MHS follows MCP’s trajectory, Anthropic’s leverage comes not from locking users into Claude but from being the most capable model behind a standard anyone can use.
Early Results from the Research Preview
Anthropic shared results from a small group of early-access partners before the public announcement.
QuEra Computing used MHS to give an AI agent control over parts of the laser system inside its neutral-atom quantum computers. Quantum computing at this level requires lasers to hold an extremely precise frequency, a process called locking. The agent developed a controller that recovers the laser’s lock 99.3 percent of the time without human intervention. Before MHS, the success rate on that recovery was 58 percent, according to The Register.
Genentech used MHS to run a drug-discovery experiment with real-time error handling. The agent managed instrument operations and responded to errors during the experiment without stopping for human review.
HHMI Janelia Research Campus, the medical research institute where MHS development began, built an agent that manages microscopes automatically, handling the kind of overnight experiment setup that typically requires trained lab technicians to stay on site.
These are not proof-of-concept demos. They are production runs in high-stakes research environments where downtime and instrument errors have direct costs.
Partner Ecosystem at Research Preview Launch
| Partner | Domain | MHS Use Case |
|---|---|---|
| HHMI Janelia | Neuroscience | Automated microscope management, real-time experiment control |
| QuEra Computing | Quantum computing | Laser stabilization, frequency lock recovery (58% to 99.3%) |
| Genentech | Drug discovery | Autonomous experiment runs with real-time error handling |
| Danaher | Life sciences | Exploring AI-enabled smart instruments and autonomous labs |
| Doosan Robotics | Industrial automation | Automated quality assurance across robotic arm fleets |
| MBF Bioscience | Neuroscience | MHS driver for ScanImage, used in hundreds of neuroscience labs |
| QIAGEN | Genomics | Instrument troubleshooting and uptime optimization on QIAsymphony |
| Tecan | Lab automation | AI agent discovery and operation of Fluent liquid handling systems |
| Universal Robots | Industrial robotics | MHS support for commercial robot arm platforms |
| Automata | Lab automation | Intelligent error handling in autonomous lab LINQ platform |
| AWS | Cloud infrastructure | MHS support via Strands Robots library for agent-to-hardware connections |
| Hugging Face | Open-source AI | Adding MHS support to LeRobot robotics library |
| Raspberry Pi | Developer hardware | MHS integration across products, Camera MHS Driver tested |
Safety Before Open Source
Anthropic is explicit that MHS will not be open sourced until the research preview generates enough safety data. The company acknowledges a gap: large language models lack physical intuition. They learned about the physical world from text and images, not from direct interaction with machines. An agent that reasons perfectly well in software can still issue a command that damages an instrument or creates a hazard in a lab.
The research preview is partly a safety evaluation exercise. Anthropic is working with its launch partners to build testing frameworks that can identify failure modes before the standard is broadly available. A physical safety roadmap is in development, separate from the software-focused safeguards policy the company has published for Claude.
Elizabeth Kelly, head of beneficial deployments at Anthropic, told CNBC: “We built this for science to sort of show the promise of AI, but there’s also huge benefits here for enterprise and for industry.”
That framing matters. Anthropic is positioning MHS as a science-first release, which allows it to build safety evidence in controlled environments before scaling to industrial settings where the consequences of agent errors are more severe and less reversible.
What This Signals About Anthropic’s Hardware Direction
MHS does not arrive in isolation. Anthropic recently hired Caitlin Kalinowski, a hardware executive with prior roles at OpenAI, Meta, and Apple, to lead its silicon team. The company is designing custom chips for its models, a move that puts it on the same path as Google (TPUs) and Amazon (Trainium and Inferentia).
The combined picture is a company that started as a pure software AI lab and is now building toward a vertical stack: models, chips, and now a protocol for connecting those models to physical infrastructure. The MCP-to-MHS trajectory mirrors how Amazon moved from software services to physical fulfillment infrastructure, or how Apple tied software performance to custom silicon.
The near-term competitive question is whether OpenAI and Google adopt MHS or build competing specs. OpenAI has invested heavily in robotics through Figure AI and its own hardware team. Google DeepMind’s Gemini Robotics platform, covered here when it launched, targets similar physical AI use cases. A fragmented standard environment, where each lab uses a different protocol, would slow the market; a single open-source standard would accelerate it.
The analogy to MCP is instructive: within 18 months of its release, MCP became a default assumption in agentic architecture. Whether MHS follows the same arc depends on how many hardware manufacturers ship drivers before competitors propose alternatives.
What Enterprise Teams Should Do Now
For the majority of enterprise teams, MHS is not an immediate procurement decision. The research preview is closed and focused on scientific research and industrial manufacturing.
The actionable steps for enterprise leaders today are:
- Map your physical workflows. Identify any process that currently requires a specialist engineer to operate a machine or interpret instrument outputs. MHS makes every such workflow a candidate for eventual agent automation.
- Track the hardware vendor list. If your supply chain or R&D stack includes instruments from Tecan, Danaher, QIAGEN, Universal Robots, or Doosan, watch for MHS driver releases. These manufacturers are already building support.
- Stay aligned on MCP. MHS extends MCP. Enterprise teams already building MCP-based integrations will be able to extend those agents to physical hardware without architectural changes once MHS is generally available.
- Watch the open-source date. Anthropic has not set a timeline, but the open-source release will be the inflection point for wide adoption. It will also include safety findings from the research preview, which will be the clearest signal of where physical AI agents can and cannot be trusted to operate independently.
For teams already building agentic workflows, the Computer Use GA release from Anthropic in August extended Claude’s reach from text to screen-based software interfaces. MHS is the next step: from screen to machine. The trajectory from language model to software agent to physical-world operator is becoming less theoretical with each release.
If you want to evaluate how physical AI fits your enterprise roadmap, talk to the Enera team.