Previewing the Model Hardware Standard

Anthropic Previews Hardware Standard to Connect AI Agents with Physical Machines

Anthropic has opened a research preview for its Model Hardware Standard (MHS), a new specification designed to allow AI agents to safely operate physical devices in labs and manufacturing facilities. Developed initially with the HHMI Janelia Research Campus, MHS aims to solve a long-standing integration problem where connecting disparate pieces of equipment can take weeks or months. By standardizing communication, the framework enables agents to control instruments like microscopes and robotic arms in parallel, reducing setup time to minutes and enabling autonomous, round-the-clock workflows.

How the Model Hardware Standard Works

MHS functions as a universal translator for hardware, using a standardized driver with simple commands—or primitives—such as “read” and “write.” This driver allows any compatible device to be discoverable on a network and communicate its capabilities and safety limits to an AI agent. Instead of relying on paper manuals or tacit knowledge, users can input machine characteristics in natural language, which the driver uses to create a reference file. An agent like Claude can then use this file to operate the device through mechanisms like the Model Context Protocol (MCP), a command line interface, or code files. This allows the agent to not only sequence steps across multiple instruments but also to observe outcomes, adjust parameters in real time, and even write deterministic scripts to execute complex tasks more efficiently.

• Standardized Driver: Uses simple primitives that any hardware device can understand.
• Device Discovery: Allows devices and agents to find and communicate with each other across networks without custom integrations.
• Natural Language Tags: Captures physical information about hardware (e.g., robot arm weight) to ensure safe operation.
• Model-Agnostic: Works with any agent harness that can access it using standard protocols like MCP.

Ecosystem Adoption Signals Broad Industry Interest

The research preview includes a number of scientific and corporate partners who have already demonstrated notable results. Researchers at Carnegie Mellon University used MHS to run dose-response experiments three times faster, while QuEra Computing leveraged it to build an agent that could recover a quantum computer's laser lock with 99.3% reliability. This early traction is bolstered by a growing list of industry adopters building MHS support into their platforms, including Amazon Web Services (AWS), Universal Robots, Doosan Robotics, Hugging Face, and Raspberry Pi. By involving these partners ahead of a wider open-source release, Anthropic is building a coalition to establish safety evaluations and best practices for AI's expanding role in controlling physical systems.

By creating a common language for AI to interact with physical hardware, Anthropic's MHS positions its models as a central orchestration layer for industrial and scientific automation, moving beyond text and image generation into real-world robotics and instrumentation.