What's this about?
Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification that lets AI agents operate physical laboratory and manufacturing equipment — microscopes, liquid handlers, robotic arms — in parallel. It began as a collaboration with HHMI's Janelia Research Campus. The pitch is integration time: getting a lab's instruments talking to each other currently takes weeks or months of bespoke work, and MHS claims to reduce that to hours. The preview is limited to a first group of research labs and advanced manufacturers.
What is a "standard" doing here, and why not just write drivers?
A standard is an agreement about how to say things, not about how to do them. Lab hardware today speaks a zoo of incompatible dialects — SCPI over serial, VISA over USB, various proprietary protocols — so every lab writes its own glue. MHS proposes one vocabulary the model already knows, and pushes the messy translation down into per-device adapters. It is the same shape as MCP, which did this for software tools; the open question is whether hardware manufacturers will adopt someone else's vocabulary.
What's the community saying?
- The single most upvoted reaction is confusion: "did they invent Bluetooth for AI? I read it twice and still don't get it."
- The clarifying answer that the thread converges on is "MCP for robotics", followed immediately by scepticism that hardware makers will play along — one commenter argues it would be easier to make robots use the existing human interfaces.
- A concrete objection: this reads like gRPC calls that MCP could already wrap, so what does MHS add?
- The strongest defence comes from someone working with lab hardware: the value isn't the transport, it's the guarantee that a given instrument is *meant* to be driven by an agent, in a field full of proprietary software and deliberate barriers.
- A researcher lists the graveyard this is entering — EPICS, TANGO, Bluesky, QCodes, Sardana, yaq — noting labs love inventing their own version of exactly this.