Anthropic published the driver language that lets Claude run lab equipment

Lab instruments each speak their own proprietary dialect. Wiring a microscope to a robotic arm or a liquid handler has historically taken specialists weeks or months of custom code, because every device describes itself differently and the connections between them are hand-rolled. Anthropic's new Model Hardware Standard collapses that into a common driver language and is handing it to outside labs and manufacturers as a research preview. The standard is the product: every programmable device that adopts the dialect becomes a Claude-reachable instrument, and any lab that builds on it is committing to whatever Anthropic does with the spec next.

The standard works by giving every device a short, plain description of itself. A liquid handler lists its channels, aspiration rate, and safety limits; a robotic arm lists its payload and workspace; a plate reader lists the wavelengths it can read. Once a device can speak that simple "read" and "write" vocabulary, a Claude agent can find it, understand what it can do, and operate it without a human writing glue code for the connection. Anthropic says this drops hardware integration time from weeks or months to hours or minutes. The same driver runs from the Model Context Protocol, a command line, or plain code, and is model-agnostic, so any agent harness can use it.

Genentech used MHS as a proof of concept to automate a BCA protein assay across a liquid handler, a robotic arm, and a plate reader, a three-instrument handoff that normally sits between a biologist and a software engineer. At the University of Washington, PhD student Zihao Song's group in the Baker and Pinglay labs built an agent-supervised qPCR pipeline with collision-free plate handoffs between a robotic arm and a liquid handler. Carnegie Mellon researchers ran serial dilution dose-response experiments roughly three times faster than before, with an AI agent orchestrating a liquid handler end to end. Anthropic's own walkthrough had Claude iteratively nudge a laser, watch the result through a camera, and package the learned sequence into a single deterministic command.

In a separate anecdote reported by Singularity Hub from a Financial Times interview, Anthropic scientist Alek Kemeny described Claude steering a microscope's mirrors and lasers in a live neuroscience experiment to locate a specific, unfamiliar structure in brain tissue, with the observing neuroscientist confirming the find. The capability is a single example attributed to one scientist through a secondary outlet, and the work was co-developed with HHMI Janelia Research Campus, not independently benchmarked.

The standard is Anthropic's, so adopting it is a vendor choice with the usual lock-in questions: any lab that builds on MHS is betting that Anthropic keeps the spec open and that competing instruments will be willing to speak the same dialect. Access today is a gated research preview, not general availability, so the weeks-to-minutes figure is a capability claim rather than a current everyday reality. On IBM's Mixture of Experts podcast, principal research scientist Kaoutar El Maghraoui called the work an "impressive proof of concept" and pointed at the other side of the same coin: in the physical world, "small errors can matter here." A robot arm that can be told what to do by an AI agent is also a robot arm that can be told the wrong thing by one.

The driver removes integration work that is currently a paid role. The lab-software specialist who used to write the glue code is the person most directly affected, and Anthropic says the standard will be open-sourced after safety work with the partner cohort. Until then, the practical question for any lab or manufacturer is the same one they would ask about any other vendor-controlled protocol: what does it cost to leave, who else is on the same dialect, and how much of "autonomous" depends on the device descriptions someone on their own side has to keep up to date.