Anthropic Previews a Physical-Device Protocol Built on MCP: Same Pattern, Different Endpoints

TL;DR: Anthropic's Model Hardware Standard, previewed August 27, extends the MCP pattern to physical instruments — same tool-call framing, new endpoint class; invite-only with research labs for now.

The research preview opens MHS to a first cohort of scientific research labs and advanced manufacturers. Initial launch partner: HHMI Janelia Research Campus. AWS Strands Robots will support MHS for connecting agents to physical devices.

What it is technically. MHS is model-agnostic: any device with a programmable interface can expose an MHS endpoint, and any MCP-compatible agent harness can discover and call it. An agent that can already write MCP servers can apply the same patterns to a robotic arm or a liquid handler — the specification handles device discovery and capability advertisement. Anthropic reports integration time dropping from weeks to hours for supported hardware.

Why it matters: If your work never involves lab automation, file this as signal that the agent-control protocol is expanding beyond software APIs to physical effectors — the same mental model as MCP, a different endpoint class. If you do build agents for research or manufacturing, the standard is planned to go open-source after the preview period; safety evaluations are actively being built with launch partners.

Limits to know: Invite-only preview; no general developer access yet. Safety evaluations for physical operation are still in progress — Anthropic is explicit that these are being built during the preview, not already validated.

Sources: Anthropic MHS research preview · CNBC: Anthropic pushes into physical world · SiliconANGLE: Anthropic previews MHS