Anthropic Previews a Model Hardware Standard for AI Agents Running Physical Devices

Anthropic has published a research preview of the Model Hardware Standard (MHS), a specification for AI agents that operate physical devices. Access is limited at launch to scientific research labs and advanced manufacturers.

What it is

A standard rather than a product. The distinction matters: Anthropic is not shipping a robot or a control system, it is proposing how models and physical hardware should talk to each other.

The restricted rollout is itself informative. Software agents that make mistakes produce bad output; agents controlling physical devices produce physical consequences. Limiting early access to labs and manufacturers is what caution looks like when the failure mode leaves the screen.

The pattern

This follows the Model Context Protocol, which Anthropic released to standardise how models connect to software tools and which was subsequently adopted well beyond Anthropic's own products. MHS applies the same play to hardware.

Whoever defines the interface shapes the ecosystem built on top of it. MCP demonstrated that a standard published early and openly can become the default before competitors publish theirs.

Where it fits with everything else

Google opened public preview of Gemini Robotics ER 2 in late July with multi-robot coordination. Two frontier labs investing in embodied agents within a month of each other is not coincidence — it is a shared read that the next capability frontier is physical rather than textual.

Why it matters to our readers

Almost nobody reading this will apply for MHS access. It matters because capabilities developed for embodied agents surface in ordinary products first: spatial reasoning improves screen control, and precise device operation improves computer use. The tools you already pay for get better before any robot arrives.

It is also worth noting what a hardware standard implies about trajectory. A lab that publishes how models should drive physical devices expects models to be driving physical devices at scale — and expects that to happen soon enough to be worth standardising now.