Anthropic launches Model Hardware Standard (MHS), a preview protocol connecting AI agents to lab and industrial hardware.
The AMW Read
Anthropic extends its MCP protocol playbook from software tools to physical hardware, updating its case study with reported pilot results spanning lab automation and quantum computing that cross into the robotics/physical-AI segment.
Anthropic launches Model Hardware Standard (MHS), a preview protocol connecting AI agents to lab and industrial hardware.
Anthropic released a preview of the Model Hardware Standard (MHS), a protocol that lets AI agents issue commands to physical devices — microscopes, robotic arms, liquid-handling systems, and quantum computers — through a shared "reference file" describing each device's live state, with safety limits like max power and max range enforced at the driver level. It is Anthropic's second industry-wide protocol after the Model Context Protocol (MCP), which standardized how agents call software tools and APIs; MHS applies the same approach to hardware, a space where prior standards (SiLA, OPC UA/LADS, AnIML) failed to gain broad adoption because they tried to hard-code every device type rather than let a model absorb the diversity. Early adopters named include Danaher, QIAGEN, and Tecan in biotech instrumentation, Doosan Robotics in industrial automation, and AWS, alongside pilots at Carnegie Mellon University and quantum-computing company QuEra.
The cited results are specific: a CMU lab used MHS to coordinate liquid handlers, plate readers, cameras, and a robotic arm and cut experiment time roughly threefold. QuEra ran a four-instance Claude agent loop to rewrite the laser-relock routine for its neutral-atom quantum computer, cutting recovery time from 150 seconds at 58% success to about 6 seconds at 99.3% success across 700 blind trials, and reducing PID-tuning residual error from an expert-tuned 15.7mV to 1.55mV with zero lock loss over 19 hours of continuous testing. MHS keeps model inference and device execution separate — Claude proposes actions and can dispatch pre-packaged scripts for latency-sensitive work, while deterministic code runs locally — the design choice behind pitching this as safe for expensive, irreversible physical actions.
For lab-automation and robotics vendors, adoption from established instrument makers and Doosan lowers integration cost for pairing hardware with agentic AI, pressuring bespoke lab-automation software while opening a distribution lane for anyone building on the standard early. For investors, the QuEra and CMU figures are self-reported pilot results, not independently audited, and MHS is still a preview release — its real reach depends on how many hardware vendors beyond the named partners commit before a rival lab ships a competing interface.
