Anthropic announced a research preview of the Model Hardware Standard (MHS), a specification that lets AI agents operate physical lab and manufacturing equipment — microscopes, liquid handlers, robotic arms — through a standardized driver. MHS began as a collaboration between Alek Kemeny on Anthropic's Beneficial Deployments team and Arco Bast, a postdoctoral scientist at HHMI Janelia Research Campus, who had built a shared memory dictionary to let mismatched lab instruments communicate at memory speed; Kemeny and Bast then worked together to integrate AI models into that interface.
What MHS does
Anthropic states that connecting AI to hardware today typically takes weeks or months per device, since most devices don't communicate with each other and require specialists to build bespoke integrations. MHS is a driver — software translating between an operating system and a hardware device — built on simple primitives like "read" (e.g. "get temperature") and "write" (e.g. "set temperature") that any hardware device can understand, plus a standard discovery format so devices and agents can find each other over a network without a custom translator program. The driver also auto-generates a reference file describing a device's characteristics, safety limits, and what it can measure or adjust, built from natural-language tags a user can write directly or dictate to an agent that interviews them about their setup. Agents control hardware through MCP, a command-line interface, or code files, and can chain driver commands into deterministic scripts for tasks that need to run faster than the agent's own reasoning allows. MHS works with any device that has a programmable interface and is model-agnostic — any agent harness can access it via standard protocols like MCP.
The specific results Anthropic reports
Anthropic names three early results directly: a drug-discovery experiment with real-time error handling at Genentech; an imaging experiment at HHMI Janelia Research Campus compressed from weeks to a single day; and laser stabilization on QuEra's quantum computers improving from 58% to 99.3%.
The limitation Anthropic discloses directly
Anthropic states plainly that Claude "learns about the physical world through text and images," so its spatial and physical reasoning "still require expert oversight." The specific example given: when working with protein samples, Genentech researchers had to guide Claude to recognize that errors caused by foaming in samples were physical failures, not software bugs, correctable only through physical adjustments rather than a code fix.
Partners building support
Anthropic lists partners adding MHS support to their own products: AWS (via Strands Robots, its library for connecting agents to physical devices), Automata (LINQ lab automation), Danaher, Doosan Robotics, MBF Bioscience (ScanImage, used in laser-scanning microscopes across neuroscience labs), QIAGEN (QIAsymphony Connect), Tecan (Fluent liquid handling), and Universal Robots. Anthropic names Hugging Face (adding MHS to its LeRobot library) and Raspberry Pi (integrating MHS across products after testing a Camera MHS Driver) as early adopters for the standard's next phase, aimed at hardware developers already using Claude Code with individual devices.
What Anthropic says is still ahead
MHS is not yet open source. Anthropic states it's using the research preview to build additional safety evaluations with launch partners and is developing a "physical safety roadmap" to strengthen safeguards against misuse, and that findings from the preview will be released alongside guidance for deploying the standard safely when it does open-source it. MHS also doesn't yet work with hardware lacking a programmable interface; Anthropic says it's working with manufacturers of such devices to build MHS drivers for them.
References: Anthropic — Previewing the Model Hardware Standard, read directly in full · related coverage: Frontier Arcade: trends & predictions