Friday, Oct 30 | --:--
Back to home

Anthropic’s Model Hardware Standard Makes the Wet Lab Look Like MCP Tools

Anthropic opened a research preview of the Model Hardware Standard (MHS) — a model-agnostic driver spec, born with HHMI Janelia, so AI agents can discover and operate microscopes, liquid handlers, and robot arms via MCP, CLI, or APIs — before a later open-source release. Physical computer-use, not a new Claude SKU.

Times of AI Desk 5 min read San Francisco, CA View as Markdown
Engraved Anthropic mark over a microscope, liquid-handler, and robot-arm plate linked by MCP driver lines, security-intaglio terracotta linework. No currency-note layout.

The Model Hardware Standard is a proposed common interface between AI agents and physical equipment. Anthropic’s research preview includes device descriptions and safety limits, while retaining the need for expert oversight.

Anthropic published “Previewing the Model Hardware Standard.” MHS is a shared driver spec so agents can operate programmable lab and factory devices — microscopes, liquid handlers, robotic arms — in parallel, with natural-language device tags that auto-generate a reference file (what it measures, what can be adjusted, safety limits). Control paths: MCP, CLI, and code files. It is model-agnostic. The work started as Alek Kemeny (Beneficial Deployments) + Arco Bast (HHMI Janelia) on a multi-vendor brain-imaging rig. Research preview is open at modelhardwarestandard.com; open-source comes after safety evals. Anthropic and its partners report early experiments at Genentech (BCA protein assay), UW Baker/Pinglay (qPCR + plate handoff), CMU (~3× faster dose-response), QuEra (an agent-developed deterministic script achieved 99.3% laser-lock recovery in a later blind test, without an agent in the loop). Vendors: AWS Strands Robots, Automata LINQ, Danaher, Doosan, MBF ScanImage, QIAGEN QIAsymphony Connect, Tecan Fluent, Universal Robots; Hugging Face LeRobot and Raspberry Pi for the next phase.

What MHS is (and is not)#

LinkEmbed

Item anthropic.com/news Aug 27
Status Research preview, not GA, not yet open-source
Scope Devices with a programmable interface only
Primitives Standardized read/write (“get temperature” / “set temperature”)
Safety caveat Claude’s spatial/physical reasoning still needs expert oversight (Genentech: foaming is a physical failure, not a software bug)
Apply modelhardwarestandard.com

This is physical computer-use: a standard, not a new Claude SKU. Its focus is a shared physical-device interface.

Limits#

  • Research preview only; open-source waits on safety evals.
  • Programmable-interface devices only — no magic for dumb hardware.
  • Anthropic is developing a physical-safety roadmap and says it will release preview findings with guidance when MHS is open-sourced.

Correction (October 14, 2026): Clarified that QuEra’s 99.3% result came from a finished deterministic script in a later blind test, not a live agent recovery rate; corrected the roadmap/publication distinction.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading