Anthropic has launched a research preview of the Model Hardware Standard (MHS), an open specification designed to help AI agents safely operate multiple physical devices in parallel. This marks the company’s most concrete step to date toward moving AI agents beyond the digital realm and into operating real-world hardware across laboratories and manufacturing plants.

The Problem MHS Solves

In a typical laboratory or production line, most instruments cannot talk to one another. Each machine relies on its own proprietary programming interface, meaning integrating them requires custom, manual development. According to Anthropic, this process often takes weeks or even months for a single lab. Furthermore, even once connected, there is no standardized way to share device data with an AI agent or allow the agent to control equipment safely.

MHS addresses this challenge through a standardized driver—middleware that translates between a computer’s operating system and the hardware device. This driver uses a simple, universal set of commands across all devices, such as “read” (for example, retrieving a temperature reading) or “write” (such as setting a target temperature). As a result, devices can automatically discover one another over a network without requiring custom translation software for each pair.

Notably, the driver also helps an agent understand how to operate hardware it has never encountered before. Critical details, such as the payload capacity of a robotic arm—essential for safe operation—often reside in paper manuals or technicians’ tacit knowledge. MHS allows this information to be documented in natural language within driver tags, automatically generating reference files that detail device characteristics: what can be measured, what can be adjusted, and where the safety boundaries lie. The agent receives sufficient context to operate the device without needing to parse external documentation.

Once connected, the agent controls hardware via three mechanisms: the Model Context Protocol (MCP), command-line interfaces (CLI), and code-based APIs. It can chain together operational steps, monitor outcomes, adjust parameters in real time, and even bundle multiple commands across different instruments into a single code file to execute long-running tasks without requiring step-by-step reasoning.

Early Partner Results

Anthropic shared the standard with a select group of research laboratories and manufacturers spanning biotechnology, robotics, and quantum computing. Early results include:

  • QuEra Computing applied MHS to let an agent control the laser system in a quantum computer. The agent autonomously developed a controller to re-establish the laser’s “lock”—the ultra-precise frequency required to interact with neutral atoms—succeeding 99.3% of the time without human intervention.
  • Carnegie Mellon University used MHS to run dose-response experiments roughly 3 times faster, with an agent coordinating liquid handlers, plate readers, robotic arms, and monitoring cameras distributed across three computers with incompatible interfaces.
  • HHMI Janelia Research Campus, where MHS originated, used it to unify and run a microscopy setup that previously required seven separate software packages lacking a shared interface.
  • Genentech deployed MHS as a proof of concept to automate BCA protein assays, which require orchestrating liquid handlers, robotic arms, and plate readers.

On the hardware side, numerous vendors are building MHS support into their equipment: AWS via its Strands Robots library, Doosan Robotics for robotic arms, Tecan for its Fluent liquid handling platform, Universal Robots, QIAGEN, Danaher, and Automata. Hugging Face is also integrating MHS into its LeRobot robotics library, while Raspberry Pi is adding support across multiple hardware products.

Remaining Limitations

Anthropic is candid about the current constraints. MHS cannot yet interact with legacy equipment that lacks a programmable digital interface. In addition, the models’ spatial and physical reasoning capabilities remain limited, still requiring human expert oversight. In one notable instance, the Genentech team had to guide Claude to understand that an error caused by air bubbles in a sample was a physical issue rather than a software bug. The standard is currently in a partner research preview ahead of an open-source release.

MHS represents a critical building block in the accelerating race toward laboratory automation across biotechnology and materials science. If widely adopted, the standard could substantially compress the time required to advance scientific research from hypothesis to result. For Anthropic, it also marks an expansion from providing frontier models to shaping the foundational operating infrastructure for AI in the physical world, echoing the role MCP has begun to play in software.