Every post we have written about MCP so far has been about software: tools, knowledge bases, OAuth, session handling. On 28 August, Anthropic opened a research preview of something in a different category entirely — the Model Hardware Standard, a specification for letting AI agents discover, operate and troubleshoot physical equipment. Microscopes, liquid handlers, robotic arms.

What the standard actually does

The Model Hardware Standard gives a device a common driver interface and a natural-language description an agent can read, the same idea an MCP tool description already uses for an API. Anthropic says it works with any device that exposes a programmable control surface, is model-agnostic, and is built to be reachable through standard protocols — MCP among them — so an agent harness does not need a bespoke integration per instrument.

Where it came from and who is testing it

  • Development started as a collaboration between Anthropic and the HHMI Janelia Research Campus.
  • The research preview is limited for now to a first group of scientific research labs and advanced manufacturers, not a general release.
  • Anthropic says early use includes automating a protein-assay procedure at Genentech and running drug-discovery experiments at Carnegie Mellon University.
  • Doosan Robotics, Tecan, Universal Robots, Hugging Face and Raspberry Pi are named as partners building or testing integrations.
  • The pitch is integration time cut from weeks or months down to hours — the same argument made for MCP itself when it replaced bespoke API integrations with one description an agent can read.

Why this belongs in an MCP roundup and not a robotics one

We wrote in June about MCP's move to the Agentic AI Foundation, and the argument then was that governance is what makes a protocol a safe bet: not owned by one vendor, moving through a public process, a real deprecation window. This is a different kind of evidence for the same bet. A protocol designed for one thing — an agent calling a software tool — turning out to describe a robotic arm with the same shape of interface is a sign the abstraction was right, not just convenient.

What this does not mean for a business chatbot

Nothing about the Model Hardware Standard is relevant to a support agent, a lead-qualification flow or a knowledge base — it is a research preview scoped to labs and factory floors, and SuperCognit's own MCP servers stay exactly what they are: HTTP tools, knowledge, and the workspace resources a business actually needs an agent to reach. We are not adding a hardware angle because one exists in a lab somewhere. The reason to notice this at all is narrower and more useful: the same open interface layer we built our MCP product on keeps getting reused for things nobody designed it for on day one, which is a better signal than any single feature release.