Anthropic has unveiled an experimental standard called MHS, which gives AI agents direct access to physical hardware. According to the company, the standard creates a unified interface and data exchange format that allows heterogeneous devices to communicate over a network without the need for individual translator programs. This moves autonomous models from a virtual environment into a real laboratory, where they can independently conduct experiments, operate instruments, and adjust parameters in real time.
What is MHS and how it works
The standard is built on the idea of unification: instead of developing a specific driver for each device, MHS defines a common data exchange protocol. A device describes its capabilities and limitations in a standardized format, after which it becomes "readable" to a language model. This approach removes one of the main barriers to automation — the need to write complex software for every new instrument.
Language instead of code: integration with Model Context Protocol
Combining MHS with the Model Context Protocol allows scientists to issue commands to equipment in plain natural language. The AI model, in turn, analyzes the current situation, adjusts parameters in real time, and resolves hardware malfunctions. In effect, the human acts as the operator-client, while routine control and monitoring are handed over to the model.
Demonstrations: calibration, microscope, and robotic arm
During the demonstrations, a Claude-level model independently calibrated a laser, verifying the result through a separate camera and fine-tuning the optical system. In another scenario, the AI focused a microscope, analyzed the image, identified an area of interest, and independently shifted the objective lens to continue the analysis. Particularly notable was the episode with the robotic arm: the model figured out how to lift an aluminum can without step-by-step algorithms from a human.
Standardized tags and safety
Because the models were trained primarily in a virtual environment, MHS includes a system of standardized tags describing the physical limitations of the "hardware." This allows the AI to instantly read the capabilities of an unfamiliar device. At the research preview stage, the company is focusing on safety assessment and the development of best practices for AI working with real equipment.
Partners and open-source plans
In the research preview, Anthropic is collaborating with a number of laboratories and manufacturers, including Amazon Web Services (Strands Robots), Hugging Face (LeRobot), Raspberry Pi, Automata, and Universal Robots. In the future, the company plans to make MHS a fully open standard (open-source), independent of specific AI model developers, which could accelerate the automation of scientific activity as a whole.
