Anthropic Launches Model Hardware Standard to Let AI Agents Control Lab Robots and Machines

Anthropic has introduced the Model Hardware Standard (MHS), a specification designed to give AI agents a unified way to read data from and control physical devices such as robotic arms, microscopes, and liquid handlers in labs and factories. The company is releasing MHS as a research preview to select labs and manufacturers, with a full open-source release planned later.

Anthropic developed the standard with HHMI Janelia Research Campus. The approach extends the logic of Anthropic's Model Context Protocol (MCP), which standardized how AI models connect to software tools and data sources, into the physical world.

How it works

Under MHS, each device gets a driver that exposes its basic read/write functions in a common format, making it discoverable by any compatible AI agent. Drivers can also encode physical constraints that software alone can't capture — such as a robotic arm's weight limits or safety boundaries — added in natural language and converted into a machine-readable reference file. Once a driver exists for a device, Anthropic says it can be reused across setups rather than rebuilt for every device-controller combination. According to Anthropic, MHS is model-agnostic and works with any device that has a programmable interface, and agents can save multi-device workflows as standalone scripts that run without a language model in the loop.

Partner test results

Anthropic cites three unverified partner case studies:

• Genentech: Claude coordinated a liquid handler, robotic arm, and plate reader to automate a protein assay. The agent optimized pipetting parameters independently but repeatedly misdiagnosed bubble-formation errors as software issues rather than physical ones, requiring human intervention to correct.
• Carnegie Mellon University: Researchers connected four device types across three computers with incompatible interfaces. Anthropic says the drivers and orchestration layer took roughly eight hours to build, versus the multiple weeks a typical vendor setup requires. Claude also autonomously adjusted a substance concentration after detecting a poor measurement curve on the first run.
• QuEra: Claude used MHS across hundreds of automated runs to build a control script that restabilizes a laser after disruptions, achieving 695 successes out of 700 attempts (99.3%) in a blind test, running fully autonomously without the language model active.

None of these results have been independently verified.

Limits and oversight

Anthropic acknowledges that Claude's physical and spatial reasoning — learned from text and images rather than direct physical interaction — remains limited. In the Genentech case, the agent needed explicit human explanation to recognize a physical failure mode. Anthropic says it is building additional safety evaluations with partners and developing a physical safety roadmap during the preview period.

Companies reportedly building or testing MHS support include AWS, Doosan Robotics, QIAGEN, Tecan, Universal Robots, Hugging Face, and Raspberry Pi.

What this means

MHS is Anthropic's attempt to replicate MCP's adoption pattern — a thin, reusable interface layer — for physical infrastructure, where integration costs have historically been a bigger barrier than model capability. The claimed integration-time reductions and the QuEra success rate are notable, but all figures come from Anthropic-selected partner examples, not independent benchmarks. The real test is whether third-party manufacturers write and maintain drivers at scale once the open-source release ships, and whether the physical-reasoning gaps seen at Genentech prove manageable outside controlled pilot environments.