Anthropic Is Developing a Standard to Connect AI Agents With Physical Equipment

Model Hardware Standard aims to connect AI agents with microscopes and robots through common drivers. The project remains in a closed research preview.

Anthropic has opened a research preview of Model Hardware Standard, or MHS, a specification designed to make it easier for AI agents to operate physical equipment. Developed with the Howard Hughes Medical Institute’s Janelia Research Campus, the project initially targets scientific laboratories and advanced manufacturing facilities.

Microscopes, robotic arms, liquid-handling systems, and optical instruments generally use interfaces specific to each manufacturer. Connecting them within a single environment therefore requires custom code that is difficult to reuse with another machine. Anthropic says MHS could reduce some integrations from several weeks to just a few hours.

The standard introduces a common driver between the agent and the equipment. This driver translates machine-specific commands into basic operations, such as reading a measurement or changing a setting. It also supplies a usable description of the device, including its capabilities, controls, measurable outputs, and the physical limits the software must follow.

This description should help an agent discover available equipment without prior knowledge of each model. A user could ask it to calibrate a laser, observe the result through a camera, and gradually correct the alignment. Anthropic says it tested this scenario before converting the resulting sequence into a deterministic script, which is better suited to fast or repetitive operations.

MHS is not a new version of Model Context Protocol. MCP is one of the channels through which an agent can access compatible equipment, alongside command-line interfaces and APIs. MHS instead describes the machine, its commands, and its operating constraints. Anthropic also presents the standard as model-agnostic and compatible with different agent frameworks.

Several manufacturers and software companies are participating in the preview, including Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, and Universal Robots. AWS is also preparing a private integration with Strands Robots. Hugging Face plans to add MHS support to LeRobot, while Raspberry Pi has tested an initial driver for its cameras. These announcements describe ongoing work rather than widespread production deployments.

The project still has several major limitations. Its code is not yet public, and access requires an application through the MHS website. The system only works with devices that have programmable interfaces. Each manufacturer must also develop or validate a driver for its hardware.

Physical safety remains a work in progress. Anthropic acknowledges that Claude still struggles with spatial and physical reasoning, making expert oversight necessary. During an experiment with Genentech, the model interpreted failures caused by foaming inside laboratory equipment as software issues. Human specialists had to step in and correct its diagnosis.

Anthropic is still developing its evaluations and physical safety roadmap. MHS should therefore be viewed as an interoperability proposal under testing, not as a certification for autonomous agents to operate any machine. Its impact will largely depend on adoption among manufacturers, the quality of individual drivers, and the strength of safeguards built into the equipment.