Anthropic has introduced a protocol designed to allow artificial intelligence agents to safely operate physical devices, an initiative aimed at bridging the gap between digital intelligence and the physical world. The Model Hardware Standard (MHS), currently in a research preview, builds on the concept of the Model Context Protocol (MCP) but extends it to laboratory equipment, factory machinery, and robotics.
The company envisions developers being able to easily connect AI models to complex hardware systems. In a illustrative example, the technology could be used to diagnose why an uranium enrichment centrifuge spun out of control. However, the practical application is already underway in research environments, notably at the Howard Hughes Medical Institute’s Janelia Research Campus, where MHS has been tested. Other organizations with suitable lab and industrial equipment can now apply to participate in the preview.
Integrating hardware with AI typically requires substantial technical overhead, often taking labs or manufacturing facilities weeks or months due to the diverse and proprietary nature of industrial interfaces. Anthropic’s MHS aims to eliminate this bottleneck, reducing setup and integration times to mere hours or minutes.
The protocol acts as a universal translation layer for hardware. Its driver software utilizes a minimal set of commands, analogous to how simple command-line tools like Bash empower AI agents, making connected devices discoverable in a standard format. Through conversational AI setups, users can tag devices to define their functions, generating a reference file that details the hardware’s characteristics for the AI. Once connected, agents can interact with machines via three pathways: MCP, command-line interfaces, and Application Programming Interfaces (APIs).
Early adopters have reported significant efficiency gains. Biotechnology firm Genentech utilized MHS to automate a drug-discovery experiment, incorporating robust error handling. Quantum computing company QuEra applied the standard to enhance laser stabilization, boosting performance from 58 percent accuracy to 99.3 percent.
The potential of the open standard has attracted a wide range of industrial partners. Amazon Web Services (AWS) plans to integrate MHS into its Strands Robots library, while Automata intends to incorporate it into its LINQ lab automation platform. Additional support is expected from companies such as Danaher, Doosan Robotics, MBF Bioscience, Qiagen, Tecan, and Universal Robotics.
Despite the enthusiasm, Anthropic acknowledges that large language models still lack intuitive physics, having learned about the physical world primarily through text and visual data. Consequently, the research preview will serve as a platform to develop safety evaluations and protective measures for AI operating in physical spaces. Let the experiments begin.