Anthropic's new Model Hardware Standard lets an AI agent run a lab's microscopes and robotic arms directly

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a shared specification that lets an AI agent operate lab and manufacturing hardware -- microscopes, liquid handlers, robotic arms, plate readers, thermocyclers, centrifuges, laser systems -- through the same simple interface, without a specialist writing bespoke integration code for every device. The standard, announced August 27, was built with HHMI Janelia Research Campus, where postdoctoral scientist Arco Bast had already built a shared memory dictionary for lab instruments; Anthropic's Beneficial Deployments team, led by Alek Kemeny, connected AI models to that interface.

MHS works through plain read and write commands that any device with a programmable interface can expose, reachable via the Model Context Protocol, a command line, or ordinary API code. State lives in a shared memory dictionary any connected program can read, which lets an agent chain a long sequence of driver commands -- run a dilution, wait, read a plate, adjust -- without re-reasoning from scratch at every step. It is explicitly model-agnostic -- Anthropic says it runs against Claude, OpenAI's models, or open-source alternatives -- and Jonah Cool, who leads Anthropic's partnerships and deployment work, pitched it around avoiding lock-in: MHS gives scientists “a standardized, easily programmable interface” to their own equipment rather than a proprietary one tied to a single vendor.

The problem MHS is aimed at is mundane and, by Anthropic's own account, the actual bottleneck in most labs: most instruments simply don't talk to each other. A microscope, a liquid handler and a plate reader from three different manufacturers each ship with their own control software, their own file formats, and no shared way for a fourth program -- an AI agent, or anything else -- to coordinate all three. Closing that gap today takes a specialist integrator and, typically, weeks to months of bespoke work per instrument. That is the specific number MHS's early pilots were built to test. Vendors that do add native MHS support -- Automata's LINQ platform and Tecan's Fluent liquid handlers among the first -- effectively agree to expose that read/write interface out of the box, shifting the integration cost from the lab's own IT staff to the manufacturer, once, rather than to every customer who buys the device. Genentech's pilot pointed the same approach at a messier problem, autonomous liquid-flow-rate optimization: the agent tuned water dispensing to about 140 microliters per second (0.016 RMSE) and a BSA protein solution to 10 microliters per second (0.181 RMSE), both within the range Anthropic says a human operator would target. Tetsuwan Scientific ran the widest test, 9,143 individual dispenses across 300 transfer types, and found the agent's compiler model beat manufacturer accuracy specs on 31 of 45 held-out runs, by 12% overall and up to 17% on the most-replicated data.

The timing lands inside a broader physical-AI push across the industry. Hugging Face -- one of the vendors on MHS's own partner list, via its LeRobot library -- debuted its own robotics hardware the same week, and the announcement follows weeks after Nvidia's reported $12.9 billion deal for Hugging Face put a major slice of the open robotics-tooling world under one company's roof. Nvidia CEO Jensen Huang has been predicting for months that “every industrial company will become a robotics company”; Kemeny's own framing is narrower and more concrete -- he describes the goal simply as removing setup friction across an entire industry's worth of mismatched hardware, not a wholesale industrial transformation.

“In the future, scientists can buy these devices and out of the box it works. That's just the process of adopting a standard.” — Alek Kemeny, Anthropic's Beneficial Deployments team

QuEra's case is the sharpest of the pilots because the task -- recovering a quantum computer's laser lock after it drifts -- is exactly the kind of continuous, tedious, instrument-babysitting work that used to require an on-call physicist. Recovery success rose from 58% under expert manual tuning to 99.3% with the agent connected, and average recovery time fell from 150 seconds to 6 seconds; the agent's PID-servo tuning also cut residual noise from 15.7 mV to 1.55 mV. Over a 19-hour unattended test, the connected agent held the lock with zero lock losses, against 1.6 manual unlocks per hour under standard expert parameters. That is the strongest single result in Anthropic's own release, and also the one furthest from a general-purpose claim: it is one instrument, one lab, one well-understood failure mode.

Anthropic is unusually direct about where MHS doesn't yet help. Its own release says Claude “struggles with physical, chemical, and biological constraints” and lacks the spatial and physical intuition to troubleshoot failures the way a bench scientist would -- the example it gives is foam forming during a liquid-handling run, which needed a human researcher's guidance to resolve. The standard also stops at devices with a genuine programmable interface: hardware that exposes only a graphical screen isn't covered yet, which rules out a lot of older lab equipment. And some agent decisions -- Anthropic doesn't specify which -- pause for a human to confirm, which the company acknowledges can stall an experiment overnight.

For now, MHS stays behind a waitlist while Anthropic works through what it calls safety evaluations and best practices with early partners -- Automata, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots and AWS's Strands Robots among them -- before any broader or open-source release. No date is attached to that transition. The distance between a research preview run under supervision and a lab technician installing MHS unattended is, by Anthropic's own account, still substantial.(MHS is built on Anthropic's 2024 Model Context Protocol, which the company has described elsewhere as functioning like a USB standard for connecting an AI model to outside software -- and, now, to physical hardware.) If the pilot numbers hold up outside a supervised preview, the bigger claim isn't that AI agents can run one microscope well -- it's that the weeks a lab loses to integration work every time it buys new equipment stop being a fixed cost of doing science at all.