Anthropic’s Model Hardware Standard gives AI agents a shared interface for physical devices

On August 27, 2026, Anthropic introduced Model Hardware Standard (MHS), a research-preview specification intended to give AI agents a consistent way to discover and operate programmable physical hardware. Rather than writing a bespoke agent integration for every microscope, liquid handler, robotic arm or controller, device vendors can expose a standardized driver with machine-readable capabilities, natural-language metadata and safety constraints. Agent harnesses can then access those devices through surfaces including MCP, command-line interfaces or APIs. Anthropic says the standard is model-agnostic and is being piloted with research institutions, robotics systems and laboratory-equipment vendors.

Physical-agent systems currently inherit the fragmentation of the hardware they control: every device has its own SDK, command vocabulary, state model and safety assumptions. A shared interface can move some integration work from agent-specific glue code into reusable device drivers and make hardware capabilities discoverable at runtime. For builders working on lab automation, robotics or other physical AI, that could shorten integration cycles and make agent orchestration more portable. The boundary is equally important: physical actions have consequences that ordinary software tool calls do not, and Anthropic explicitly describes MHS as a research preview requiring further safety work and expert oversight rather than a mature autonomous-control standard.

The proposed standard defines a device-facing layer with read and write primitives, discoverable capabilities and metadata that can describe operating instructions, limits and safety considerations. An agent does not need to know each vendor's proprietary SDK in advance; the hardware adapter exposes a consistent surface that an agent harness can inspect and invoke.

Anthropic positions MHS below higher-level agent interfaces. A hardware driver can be exposed through MCP, a CLI or another API, allowing an existing agent system to treat physical devices as tools while MHS handles the device-specific contract. That separation is useful because it keeps model choice and agent harness independent from the hardware integration layer.

Anthropic describes work with organizations including Genentech, the University of Washington and HHMI Janelia, alongside robotics and laboratory-hardware partners. Example systems span microscopes, liquid handlers and robotic equipment. The current evidence is primarily Anthropic and partner reporting, so claims that integrations fall from weeks or months to hours or minutes should be treated as early implementation evidence rather than a general benchmark.

A device description can tell an agent about permitted actions and operating boundaries, but metadata is not a physical safety system. Anthropic says more work is needed before open-sourcing the standard and highlights limitations in model physical reasoning. Builders should retain hardware interlocks, permission boundaries, simulation or dry-run paths and human approval for consequential operations.