Mumbai: Anthropic has opened a research preview of its Model Hardware Standard (MHS), a proposed shared specification designed to allow AI agents to safely operate and coordinate physical equipment used in laboratories, manufacturing facilities and other technical environments.
The standard is being made available initially to a select group of scientific research labs and advanced manufacturers. Anthropic says MHS can allow AI agents to operate multiple devices, including microscopes, liquid handlers and robotic arms, in parallel.
The aim is to tackle a rather mundane but stubborn problem in the AI-powered physical world: getting machines from different manufacturers to communicate with one another.
Laboratories and manufacturing facilities can currently spend weeks or months integrating equipment because individual devices often rely on different programming interfaces. MHS seeks to cut that integration time to hours or even minutes by providing a common software layer between AI agents and programmable hardware.
The standard uses a driver that translates between a computer’s operating system and a physical device. It provides basic commands such as reading information from a machine or writing a new setting to it, while also making devices discoverable in a common format.
Crucially, the system is designed to give AI agents more information about the equipment they are controlling. Manufacturers or users can add natural-language information about a machine’s characteristics, capabilities and safety limits. This information is then turned into a reference file that an AI agent can use when operating the equipment.
MHS is model-agnostic and can be accessed through standard protocols such as the Model Context Protocol. Agents can use the system through MCP, command-line interfaces and application programming interfaces, allowing multiple machines to be orchestrated through a single workflow.
Anthropic says early testing showed how an AI agent could operate in a manner resembling an experimental scientist. In one example, Claude adjusted a laser, observed the resulting beam through a camera and repeatedly refined the setting. It then converted what it had learned into a deterministic script that could align the laser with a single command.
The research preview has attracted partners across biotech, robotics, quantum computing and manufacturing.
Amazon Web Services plans to support MHS through Strands Robots, while Automata is adding support to its LINQ lab automation platform.
Danaher and Anthropic are exploring applications for smart instruments and autonomous laboratories. Doosan Robotics is testing MHS with robotic arms for automated quality assurance and coordination across multiple robots.
Meanwhile, MBF Bioscience is developing an MHS driver for ScanImage, its software for laser-scanning microscopes. QIAGEN has built a proof of concept for its QIAsymphony Connect platform, exploring how AI agents could help troubleshoot equipment and guide operators through recovery.
Tecan is adding MHS support to its Fluent liquid-handling platforms, while Universal Robots plans to add support to its robotics platform.
Anthropic says Hugging Face is adding MHS support to LeRobot, its robotics library, while Raspberry Pi is enabling integration across several products following tests with its Camera MHS Driver.
Anthropic is positioning the preview as a testing ground before making MHS open source. The company acknowledges that AI systems still have limitations when reasoning about the physical world and will require expert oversight.
In one example involving protein samples, researchers at Genentech had to guide Claude to recognise that foaming was a physical problem rather than a software error. Such cases highlight the gap between an AI agent’s ability to reason from text and images and the practical realities of operating machinery and experiments.
MHS also currently works only with equipment that has a programmable interface. Anthropic is working with manufacturers to extend the standard to devices that do not yet offer one.
The company plans to use the research preview to develop further safety evaluations and protections for AI systems operating physical equipment. It is also working on a physical safety roadmap covering potential misuse before releasing the standard as open source.
The broader ambition is significant: rather than having AI operate one machine at a time, MHS could give agents a common language for coordinating entire laboratories and production environments. If the standard can make that vision safe and reliable, the next AI interface may not be a screen or keyboard, but the physical world itself.