The Model Hardware Standard, or MHS, is Anthropic’s new research-preview standard for letting AI agents safely operate physical equipment. Think microscopes, robotic arms, liquid handlers, plate readers, cameras, and other programmable devices used in scientific research and advanced manufacturing.

The problem MHS tries to solve is simple to state and painful in practice: lab and factory devices often speak different technical languages. In Anthropic’s launch video, researchers describe spending huge amounts of time wiring devices together, debugging hardware, and building custom software before the actual science can begin. MHS gives those machines a shared interface so an AI agent can discover what each device can do, read its state, and issue controlled commands.

According to Anthropic’s announcement, MHS uses standardized drivers with simple primitives such as read and write. Those drivers can describe a device’s characteristics, what it can measure, what it can adjust, and which safety limits should be enforced. Once connected, an agent can sequence steps across devices, monitor results, adjust parameters, and sometimes recover from hardware errors. Anthropic says MHS is model-agnostic and can work with standard protocols such as Model Context Protocol, command line tools, and APIs.

For business leaders, the important point is that MHS moves AI agents from software-only workflows toward the physical world. The early examples in Anthropic’s announcement include microscope control, robotic arm coordination, liquid handling, drug discovery experiments, qPCR monitoring, and dose-response automation. But it is not a finished public standard yet. Anthropic says MHS is in limited research preview with partners, with open source release planned later after more safety evaluations and best practices are developed.

If MHS has you thinking about the safety and accountability questions that come with AI acting in the physical world, AI Governance gives business leaders a practical grounding in responsible adoption, oversight, risk management, and decision rights.*