Anthropic PBC has launched a research preview of the Model Hardware Standard (MHS), a shared specification designed to enable artificial intelligence agents to safely and autonomously operate physical laboratory and manufacturing equipment. The company is opening MHS to a first group of scientific research labs and advanced manufacturers to collaborate on building safety evaluations and developing best practices for AI systems operating physical equipment. As Elizabeth Kelly, Anthropic's head of beneficial deployments, told CNBC, the company built the standard with science in mind first, noting there are "huge benefits here for enterprise and for industry." The standard is designed to be model-agnostic, meaning it can work with AI models beyond Anthropic's Claude, positioning Anthropic to compete for the connective layer between AI reasoning and physical execution. To apply for access to the research preview, interested organizations can visit Anthropic's website.
The MHS operates as a model-agnostic interface compatible with any device featuring a programmable connection, introducing a common set of operational primitives such as read and write commands. A core innovation is its metadata tagging system, which captures tacit operational knowledge typically found in physical manuals and automatically formats it into reference files that equip AI agents with context needed to discover, understand, and safely control unfamiliar equipment. Agents can now sequence multi-instrument tasks, monitor outputs in real time, adjust parameters dynamically, and execute long-running operations via chained command scripts to bypass reasoning latency. The standard embeds operational constraints such as speed and angle limits to prevent unsafe device operations, ensuring AI control remains within safe parameters. Notably, MHS allows hardware to communicate seamlessly with AI systems through established protocols like the Model Context Protocol, creating a unified communication framework across different AI systems. MHS provides a common driver interface, natural-language device tags that let a model describe and locate hardware in plain language, and built-in compatibility with Anthropic's existing Model Context Protocol, extending the same communication logic into the physical world.
Early testing results demonstrate MHS's practical impact on operational efficiency across multiple institutions. At Carnegie Mellon University, researchers used MHS to connect laboratory equipment spread across three computers with incompatible interfaces, completing the system build in roughly eight hours compared to several weeks for a typical vendor-built setup. An AI agent then ran a dose-response experiment, rejected an inadequate result, changed the concentration range and repeated the experiment without human intervention, with researchers reporting the experiments ran approximately three times faster than before. At the University of Washington, researchers connected six instruments in less than a week, using an AI agent to monitor experiments remotely and coordinate a robotic arm with a liquid handler for automated sample transfers. At Genentech, researchers connected Claude through MHS to a liquid handler, robotic arm, and plate reader for a protein assay, with Claude running experiments, analyzing results, and adjusting liquid-transfer parameters while recovering autonomously from some equipment failures. However, researchers found the model still struggled with physical phenomena such as bubbles in liquids, exposing a significant limitation of current AI reasoning capabilities. According to Financial Times, scientists have observed AI agents tracking, adjusting and repeating experiments in ways similar to human researchers, with one neuroscience experiment showing Claude examining live brain tissue and identifying the alveus, a specific part of the brain.
The MHS framework extends beyond laboratory automation to address industrial manufacturing challenges. A modern factory contains many programmable machines including robots, cameras, sensors, and production equipment, but these systems are often built as separate components with software written specifically to make them work together. MHS enables an AI agent to receive high-level instructions to inspect a batch, coordinate cameras and sensors, instruct robots to move suspect components, and adjust machine parameters based on inspection results. The agent operates at a higher level, deciding how different pieces fit together while individual machines perform physical work. Anthropic says MHS agents can sequence steps across instruments, monitor results, and adjust parameters as conditions change, allowing AI models to avoid sitting in the control loop for every millisecond of robot movement. For long-running operations, agents can chain driver commands into code so hardware executes sequences without requiring the model to reason through every individual step. This approach means AI models don't need to handle every millisecond of robot movement, enabling the low-level operation to remain deterministic while AI operates at higher levels, handling planning, monitoring, and changes when unexpected situations occur.
Anthropic has been testing the hardware framework with multiple industry stakeholders including Amazon Web Services, Danaher, Doosan Robotics, Hugging Face Inc., and Raspberry Pi. Amazon Web Services is integrating MHS through its Strands Robots library, while Danaher and Doosan Robotics are testing the standard across smart instruments and automated arms. Hardware developers at QIAGEN, Tecan, and Universal Robots are building direct MHS compatibility into their platforms, with independent software initiatives led by Hugging Face and Raspberry Pi embedding the standard into their robotics and embedded systems toolchains. Safety evaluations are currently being conducted with a handpicked group of partners including HHMI Janelia, Genentech, and Carnegie Mellon University, with Anthropic planning to open-source the standard following the same path it took with the Model Context Protocol in 2024. The company isn't releasing MHS to the general public yet, instead making it available as a research preview to select organizations in science, robotics, and manufacturing, with plans to eventually open-source it so any device manufacturer in any industry could adopt it. According to Financial Times, Anthropic plans to open-source MHS but will first work on safeguards to reduce the risk of errors when AI agents interact with physical equipment.
