Anthropic wants AI agents to safely run lab equipment, and is testing a shared standard for it
The Model Hardware Standard is a research preview aimed at microscopes, liquid handlers, robotic arms, and other programmable lab and manufacturing instruments -- model-agnostic, built to work alongside MCP, with a first cohort that includes Genentech, Carnegie Mellon, and HHMI Janelia.
Anthropic announced the Model Hardware Standard on August 27, 2026, describing it as "a shared specification for AI agents to safely operate physical devices." It targets the instruments that fill a research lab or manufacturing floor -- microscopes, liquid handlers, robotic arms, plate readers, qPCR machines, laser systems -- and, per Anthropic, works with "any device that has a programmable interface," not just equipment Anthropic itself builds toward.
Model-agnostic, and built to sit alongside MCP
Anthropic is explicit that MHS isn't Claude-exclusive: it's "model-agnostic, and any agent harness can access it using standard protocols, such as the Model Context Protocol," meaning a Claude, GPT, or Gemini agent could in principle drive the same instrument through the same interface. The company is sharing an early version with a first cohort -- Genentech, Carnegie Mellon, HHMI Janelia, QuEra, the University of Washington, and Tetsuwan Scientific among them, per Anthropic's own announcement, with Fortune separately reporting additional early partners including Universal Robots, AWS, Doosan Robotics, Danaher, and Hugging Face -- ahead of a planned open-source release once safety evaluations are further along.
A harder version of the same boundary problem
This site has spent the past week on what happens when the boundary between an agent's reasoning and its access to a system isn't actually verified -- OpenAI's and Anthropic's own testing incidents, both covered here, were failures of exactly that boundary in software. MHS raises the stakes on the same question by extending it to physical actuation: a standardized interface that makes it faster for an agent to drive a robotic arm or a laser system is also, by design, removing friction from the same class of action whose software equivalent this site has spent the past week writing about. A shared, well-documented interface is a better place to build safety controls than a dozen bespoke ones -- but it's still only as safe as whatever verifies, at runtime, that the agent issuing a command through it is authorized to issue that specific command to that specific device.