Anthropic previews a common interface for AI agents to operate lab hardware
Anthropic has opened a research preview of the Model Hardware Standard, or MHS, a proposed common interface for connecting AI agents to programmable equipment. The first participants are scientific laboratories and advanced manufacturers, where a single workflow may need to coordinate microscopes, liquid handlers, cameras, robotic arms and other instruments built around incompatible controls.
The practical pitch is less dramatic than “AI runs the factory,” but potentially more useful: give devices a consistent way to describe what they can do, expose controls and share measurements. An agent could then supervise a multi-instrument process without engineers writing a new translator for every connection.
What the standard actually changes
Most laboratory and industrial automation still depends on vendor-specific software, custom scripts and specialist knowledge about each machine. MHS introduces a standardized software driver with basic operations such as reading a measurement or writing a setting. It also makes connected devices discoverable across a network in a common format.
The driver can carry natural-language information that code alone may not reveal, including a machine’s physical characteristics, available measurements, adjustable parameters and enforced safety limits. That reference information is intended to help an agent understand unfamiliar equipment before it issues commands.
Anthropic says MHS is model-agnostic and can be accessed through the Model Context Protocol, a command-line interface or APIs. The common interface does not require an AI model: conventional software can use the same drivers. The agent layer becomes relevant when a workflow must sequence steps, monitor results and change parameters as conditions develop.
Why laboratories are the first test
Research labs make a demanding proving ground because experiments routinely combine specialized instruments from multiple suppliers. Anthropic says the project began with HHMI Janelia Research Campus, where a brain-imaging setup brought lasers, cameras and motorized components together through a shared interface.
The company’s announcement describes early work in microscopy, biotechnology, robotics and quantum computing. One Carnegie Mellon University example used drivers for several instruments plus an orchestration layer to run a protocol; Anthropic reports that setup took about eight hours, compared with several weeks for a typical vendor-built integration. That is a partner case study, not an independently benchmarked industry result, so it should be treated as an indication of potential rather than a general performance guarantee.
CNBC independently confirmed that access is limited to a selected group in science, robotics and manufacturing. Early participants named by Anthropic and reported by Ars Technica include Raspberry Pi, Hugging Face, Amazon Web Services, Automata and Universal Robots. Anthropic plans to make MHS open source after the preview, but it has not announced a public-release date.
The safety layer is the unfinished part
Standardizing access does not make physical operation safe by itself. A mistaken database query can often be reversed; a bad instruction to a robot, heater or laser can damage a sample, stop production or injure someone. Device-level limits, permissions, isolation and human stop controls therefore matter as much as the agent’s reasoning.
Anthropic acknowledges that current models have limited spatial and physical reasoning. In one example disclosed by the company, Genentech researchers had to help Claude recognize that foaming in protein samples was a physical failure rather than a software fault. The preview will be used to develop safety evaluations and deployment practices, while Anthropic says it is also preparing a physical-safety roadmap.
For prospective adopters, the important questions are operational: Which commands can the agent issue? Which limits are enforced by hardware rather than written guidance? Can every action be logged and replayed? What happens when networking, sensors or the model fail? And does an expert approve high-consequence steps? MHS may provide a shared language for these controls, but the preview has not yet established a mature, independently evaluated safety regime.
What to watch next
MHS is currently an invitation-only experiment, not a finished universal connector. It works only with equipment that already has a programmable interface, and manufacturers may still need to build and maintain drivers. Adoption will depend on whether competing vendors accept the specification, whether implementations behave consistently and whether the open-source release arrives with credible conformance and safety tests.
If those pieces come together, the near-term win could be shorter integration projects and more reusable automation—not fully autonomous laboratories. The preview is significant because it moves the AI-agent discussion from screen-based tasks to machines with real-world consequences, where interoperability and hard safety boundaries will have to advance together.
Read article 
