Explained: What is Anthropic’s Model Hardware Standard and how AI agents could start controlling real-world machines

Anthropic has opened a research preview of its Model Hardware Standard (MHS), a framework designed to let AI agents operate physical equipment used in scientific research and advanced manufacturing. Instead of researchers writing separate software to make every machine communicate with an AI system, MHS provides a common interface between AI agents and hardware.

That could allow an AI agent to control several machines, monitor what they are doing and adjust an experiment as it runs. Anthropic developed MHS with HHMI Janelia Research Campus and is initially making it available to a limited group of research labs and manufacturers. The company says it plans to open-source the standard after using the research preview to develop safety evaluations and best practices.

What is Anthropic’s Model Hardware Standard?

Scientific laboratories and advanced manufacturing facilities often use equipment from different vendors. A microscope may have one software interface, a robotic arm another and a liquid handler a third. Getting these machines to work together can require custom integrations. Anthropic says setting up and integrating hardware can take weeks or even months. MHS is intended to reduce that work to hours or minutes.

It does this by introducing a standardised driver between the computer and the physical device. The driver translates simple commands such as "read" and "write" into actions the hardware can understand.

MHS also creates a standard description of each connected device. This can include what the machine can measure, what can be adjusted and what safety limits apply. Anthropic gives the example of a robotic arm's weight, which can matter when determining how it should be moved safely. This information can be captured in natural language and turned into a reference file for the AI agent.

How does MHS let AI control multiple machines?

Once hardware has been connected, an AI agent can interact with it through three mechanisms: the Model Context Protocol (MCP), a command-line interface and code files or APIs. This allows an agent to orchestrate several machines through a common interface.

The agent can receive information from equipment, sequence actions, monitor results and change parameters as an experiment progress. For longer or repetitive operations, it can also turn what it has learned into deterministic code, allowing machines to execute a process without an AI model reasoning through every individual step.

Anthropic says it saw this during testing with a laser. Claude adjusted the laser, observed the result through a camera and repeated the process to understand how the controls affected the beam. It then converted what it learned into a script that could align the laser with a single command.

What can MHS actually do today?

The technology is still in a research preview, but Anthropic has demonstrated it across laboratories working in drug discovery, microscopy and quantum computing. At Genentech, researchers used MHS to connect a liquid handler, robotic arm and microplate reader for a proof-of-concept BCA protein assay. Claude ran the experiment, analysed the results and adjusted liquid-handling parameters. It found that water worked best at a flow rate of about 140 microlitres per second, while the more viscous protein solution required about 10 microlitres per second.

The experiment also exposed a limitation. When bubbles caused errors, Claude initially tried repeating the operation with different parameters, which made the problem worse because the additional mixing created more bubbles. Researchers had to tell the model that the error was caused by a physical problem and explain how to respond to it.

At Carnegie Mellon University, researchers used MHS to coordinate a liquid handler, plate reader, robotic arm and cameras for serial-dilution experiments. The team said it built the required drivers and orchestration layer in about eight hours, compared with several weeks for a typical vendor-built setup. The resulting experiments ran roughly three times faster.

MHS was also tested at HHMI Janelia, where researchers used it to bring together complex microscopy equipment that previously depended on seven different vendor programs.

At quantum-computing company QuEra, an AI agent was used to help recover a laser's precise frequency lock. In a blind test involving 700 induced disturbances, the resulting deterministic script successfully restored the lock 695 times, or 99.3% of the time.

Why is this important for AI?

AI systems have largely been confined to the digital world. They can write code, analyse documents and operate software, but controlling physical equipment has traditionally required another layer of specialised engineering.

MHS attempts to remove some of that barrier. If a laboratory has several programmable machines, an AI agent could potentially coordinate them without engineers having to build a new integration for every combination of devices.

The bigger implication is closed-loop experimentation. An AI agent could propose an experiment, instruct the equipment to run it, receive the results, analyse them, change the parameters and run another experiment. Instead of only analysing scientific data, the AI could participate in the physical process that generates it.

Anthropic describes a future in which scientists specify a high-level biological goal while AI agents coordinate the physical workflow, execute experiments, analyse results and use them to plan the next round. This could be particularly significant in drug discovery, where researchers need to test large numbers of molecules and experimental conditions.

Does this mean AI can run a laboratory without humans?

Not yet. The demonstrations are proofs of concept, and Anthropic acknowledges that current models have difficulty understanding physical reality. More complicated experiments will require additional optimisation and safeguards.

The company is using the research-preview phase to develop safety evaluations and determine where human approval should be required, particularly for higher-risk decisions.

MHS does not make an AI agent physically competent by itself. It gives the agent a standardised way to communicate with machines and exposes information about their capabilities and safety limits. The resulting system still depends on the AI model, hardware, drivers, safeguards and surrounding software.

What happens next?

Anthropic is treating MHS as a limited research preview and is inviting organisations across science, robotics, electronics and manufacturing to participate. The company says the standard is model-agnostic, meaning it is not intended to work only with Claude. Any agent harness should be able to access it through standard protocols such as MCP.

The company is also working with hardware and robotics organisations to expand support. Early adopters include Hugging Face, which is adding MHS support to its LeRobot robotics library, and Raspberry Pi, which is enabling MHS integration across some of its products.

Anthropic plans to open-source MHS after the research-preview phase. If the standard gains adoption, its significance could extend beyond Anthropic's own models. The more hardware that supports a common interface, the easier it could become for different AI agents to interact with physical machines.