Anthropic previews MHS, a device-driver standard for AI agents in labs and factories
- Anthropic opened a research preview of the Model Hardware Standard (MHS) to an initial group of scientific labs and advanced manufacturers, with plans to open source the specification later.
- MHS drivers standardize device discovery plus basic read and write commands, translating between AI agents and programmable equipment such as microscopes, liquid handlers, and robotic arms.
- Each driver can include natural-language tags for device characteristics and enforced safety limits, then generate a reference file describing what the device can measure or adjust and how an agent may operate it.
- Agents can control MHS devices through MCP, a command-line interface, or code APIs; they can sequence instruments, monitor data, change parameters in real time, and run chained commands without reasoning through every step.
- Anthropic says Claude tested laser alignment by adjusting a laser, evaluating camera output, then writing a deterministic script that ran the learned alignment procedure as one command.
Hacker News 의견들
I still do not get what MHS is. Did Anthropic just invent Bluetooth for AI?
I see it as MCP for robotics. I would rather have robots use existing human UIs than depend on hardware manufacturers adopting another interface.
This sounds like a basic instruction set translated by per-device drivers. The real work is still writing drivers for every piece of hardware an LLM needs to access.
Lab hardware is full of proprietary software and automation barriers. The point may be to certify that a device can intentionally interface with an agent, like MCP versus letting an agent make raw HTTP requests.
There are already many lab-control packages, including EPICS, TANGO, Bluesky, QCodes, Sardana, and yaq. Anthropic appears to be making the n+1 abstraction layer for heterogeneous protocols such as SCPI, VISA over USB, and proprietary interfaces.
This sounds a lot like PyLabRobot.
Putting LLMs in control of robots has more serious consequences than software-only automation because the model acts in the physical world.
MCP, MHS, and similar standards look like fairly obvious tool interfaces that Anthropic can use as model-training scenarios.
I think it is useful beyond AI because ordinary scripts could automate processes through the same interface. Calling every universal software interface a "Model" acronym is annoying.
Anthropic should have started with a survey of existing protocols and shown why it can be trusted to run a standard. MCP began awkwardly, even if its origin as internal tooling explains some of that.
I suspect Anthropic gets months to post-train Claude on a protocol that competitors cannot even inspect yet. Its control of Claude gives it a credible path to make that race establish the standard.
This has roughly the same goals as ROS 2.
I read it as an AI-facing hardware abstraction and device-description layer, similar to OPC UA or SiLA. MHS sits between MCP, CLI, or code and vendor APIs such as SCPI, OPC UA, and ROS, while describing capabilities, metadata, and safety limits in machine-readable form.
The idea makes sense, but a standard that is not publicly readable is not really a standard yet. USB and CAN did not require permission just to inspect the specification.
Silicon Valley seems to reinvent ROS every few years.
This is awkward beside reports that models can deceive operators or hack systems. Anthropic is proposing AI hardware interoperability while showing models controlling microscopes and other physical equipment.