Anthropic’s New Model Hardware Standard Connects AI Agents to Robots and Lab Equipment

Anthropic has introduced the Model Hardware Standard, or MHS, a shared specification designed to let AI agents safely discover, communicate with and operate programmable physical equipment. The standard is entering a research preview with a limited group of scientific labs, robotics companies and advanced manufacturers before Anthropic plans to release it as open source.

MHS was developed initially with the HHMI Janelia Research Campus and is intended to reduce the work required to connect AI systems with laboratory and manufacturing equipment. Anthropic said hardware integrations that can take weeks or months could be reduced to hours or minutes by using a common interface instead of custom software for every device.

The system is designed to work with equipment that exposes a programmable interface, including microscopes, liquid handlers and robotic arms. Once connected, AI agents can coordinate multiple devices, monitor results, adjust settings as conditions change and, in some cases, recover from hardware errors without human intervention. “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,” Elizabeth Kelly, Anthropic’s head of beneficial deployments, told CNBC.

At the center of MHS is a standardized software driver that translates between hardware and an AI system. Rather than requiring a custom integration for every instrument, the driver exposes a basic set of operations such as reading a measurement or changing a setting. The standard also gives agents structured information about the physical characteristics and operating limits of equipment. Users can add details in natural language, including information that may not be obvious from software alone, such as a robot arm’s weight or safety constraints. MHS then generates a reference file describing what the machine can measure, which settings can be changed and which limits must be enforced.

Once a device is available through MHS, agents can control it through the Model Context Protocol, a command-line interface or APIs. Those methods can be combined so a single agent can coordinate several pieces of equipment as part of one workflow.

Anthropic said Claude has already been tested on physical experiments using the standard. In one example, the model adjusted a laser, inspected the result through a camera and repeated the process to determine how its changes affected the beam. Claude then converted what it learned into a deterministic script so the alignment process could run without requiring continuous reasoning from the model.

That approach allows an agent to handle higher-level decisions while delegating repetitive or time-sensitive operations to code. It can supervise experiments, respond to changing conditions and sequence work across devices, while scripts execute lower-level instructions when continuous model involvement would be inefficient.

Anthropic has been testing MHS with companies and research organizations across biotechnology, robotics, quantum computing and manufacturing. The company said early deployments have shortened hardware integration times, accelerated experimental iteration and helped with real-time equipment operation and fault detection.

Several hardware and software companies are also adding support for the standard. Amazon Web Services plans to integrate MHS with Strands Robots, its library for connecting agents with physical devices, and will provide preview participants with an early version of the software.

Automata is adding support to its LINQ lab automation platform, while Doosan Robotics is testing MHS with robotic arms for quality assurance and coordination across multiple machines. Tecan is adding support for its Fluent liquid handling systems, and Universal Robots plans to integrate the standard into its robotics platform. Other participants include Danaher, MBF Bioscience and QIAGEN. MBF Bioscience is developing an MHS driver for ScanImage, software used with laser-scanning microscopes, while QIAGEN has built a proof of concept around its QIAsymphony Connect nucleic acid purification platform. Hugging Face is also adding MHS support to its LeRobot robotics library. Raspberry Pi is working on integrations across several products after testing a Camera MHS Driver.

Anthropic stressed that the system remains early. Claude’s understanding of physical environments is still limited by the fact that the model primarily learns from text and images, which means human expertise remains necessary in situations where physical behavior is difficult for the model to interpret.

In one example cited by the company, Genentech researchers had to help Claude understand that foaming in protein samples represented a physical problem rather than a software failure and required a physical correction.

MHS also cannot currently control devices that lack programmable interfaces. Anthropic said it is working with manufacturers to develop drivers for more hardware as part of the next phase of the project.

The company plans to use the research preview to create additional safety evaluations and deployment guidance before opening the standard to the public. Anthropic is also developing a physical safety roadmap focused on reducing misuse risks as AI systems gain greater control over real-world equipment.

Although Anthropic developed the standard, MHS is model agnostic. Any compatible agent system can use it through supported protocols, including MCP, meaning organizations are not required to use Claude.

Anthropic is now accepting applications from organizations interested in joining the research preview. The company said findings from the program will be released alongside safety guidance when MHS is eventually open-sourced.

See more here:

This analysis is based on reporting from CNBC & Anthropic.

Image courtesy of Anthropic.

This article was generated with AI assistance and reviewed for accuracy and quality.