Anthropic is pushing Claude beyond the computer screen and into the physical world with a new program designed to let AI agents operate laboratory equipment, robots and other programmable machines.
The AI company on Thursday unveiled a research preview of its Model Hardware Standard (MHS), a specification intended to give AI agents a common way to discover, communicate with and safely control physical devices. Anthropic is initially making the preview available to a group of scientific research labs and advanced manufacturers, with plans to eventually open-source the standard.
MHS is designed to allow AI agents to operate multiple machines simultaneously, including microscopes, liquid handlers and robotic arms. Anthropic said potential applications range from routine drug-discovery experiments to calibrating lasers on quantum computers. The project began as a collaboration between
Anthropic and HHMI Janelia Research Campus.
The effort targets a longstanding problem in laboratories and manufacturing facilities: getting equipment from different vendors to work together. Anthropic said integrating hardware can take weeks or months because devices often have their own programming interfaces and require specialists to build custom connections.
How It Works
MHS introduces a standardized software driver that acts as an intermediary between a computer and a physical device. It uses basic commands such as reading a machine's temperature or writing a new temperature setting, while also making devices discoverable in a common format.
The standard is also intended to give AI agents information they may need to safely operate unfamiliar equipment. Users can describe characteristics and safety limits in natural language, allowing the system to generate a reference file detailing what a machine can measure, what can be adjusted and which limits must be enforced.
Once connected, devices can be controlled through Anthropic's Model Context Protocol, a command-line interface and application programming interfaces. Agents can then coordinate multiple instruments, monitor experimental results and change parameters as conditions evolve.
Anthropic said early testing showed Claude could interact with physical experiments in an exploratory fashion. In one example, Claude adjusted a laser, used a camera to observe the resulting beam movement and repeated the process until it understood the sequence. It then converted what it learned into a deterministic script that could align the laser with a single command.
What's Next
A number of companies are already testing or adding support for MHS. Amazon Web Services plans to support the standard through Strands Robots, while Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan and Universal Robots are among the companies exploring integrations.
The potential applications extend from robotics and manufacturing to biotechnology and neuroscience. QIAGEN, for example, is testing MHS on its nucleic acid purification platform to help AI agents troubleshoot instrument problems and guide recovery, while Tecan is adding support to its Fluent liquid-handling systems.
But Anthropic acknowledged that the technology still has significant limitations. Claude's understanding of the physical world remains imperfect and requires expert oversight. In one example involving protein samples, Genentech researchers had to help Claude recognize that foaming was a physical problem rather than a software bug.
MHS also currently works only with hardware that has a programmable interface. Anthropic said it is working with manufacturers to expand compatibility and plans to use the research preview to conduct additional safety evaluations.
