Anthropic's Model Hardware Standard Lets AI Agents Run Lab Equipment Directly — Cutting Integration From Months to Minutes

Anthropic, working with HHMI Janelia Research Campus, has released a research preview of the Model Hardware Standard (MHS) — a shared specification that lets AI agents directly operate physical lab and manufacturing equipment, connecting to devices over MCP, a CLI, or generated code, with standardized device discovery and natural-language tags replacing equipment manuals. Anthropic says it cuts hardware-integration time from weeks or months down to hours or minutes. Six partners piloted it with concrete results: Genentech automated BCA protein assays across liquid handlers and plate readers, with Claude auto-optimizing flow rates; Carnegie Mellon ran dose-response experiments three times faster; and QuEra Computing’s laser-locking recovery success rate jumped from 58% with prior methods to 99.3%. HHMI Janelia used it to unify microscopy rigs that previously required seven separate vendor programs. Anthropic is sharing the early version with select scientific and manufacturing partners to build out safety evaluations before eventually open-sourcing it.

The MCP dependency isn’t incidental — it’s the same protocol Anthropic has spent the past several weeks scaling for enterprise software integration. In a late-July update, Anthropic shipped MCP’s fifth major spec update, moving the protocol to a stateless request/response core so servers can run on serverless infrastructure, and reported MCP had passed 400 million monthly SDK downloads, a 4x year-over-year increase, with Claude’s connector directory listing more than 950 servers. MHS is effectively that same integration pattern extended from software APIs to physical hardware. It’s also not the only lab pushing AI into scientific infrastructure: Meta’s open-source SAM 3 and DINOv3 vision models are already running the SYNAPS-I project inside the U.S. Genesis Mission at Department of Energy national labs, processing tens of petabytes of imaging data a year — in one demonstration, compressing a month of manual grapevine drought-resilience analysis into roughly 15 minutes on 300 A100 GPUs, entirely within an air-gapped environment. Between MHS and SYNAPS-I, agentic AI’s next integration frontier looks less like a new chat interface and more like direct control of the equipment scientists already have on the bench.