Anthropic released a research preview of the Model Hardware Standard (MHS) on August 27, a communication spec that lets AI agents control physical equipment directly, spanning microscopes, liquid handlers, robotic arms, centrifuges, spectrometers, incubators and other lab and advanced-manufacturing hardware. The headline number is onboarding cost: wiring a single instrument into an automated workflow used to take weeks to months; with MHS, Anthropic says it now takes hours, sometimes minutes.
The approach isn't exotic. At the driver level, MHS abstracts device capabilities into two primitives — read, to pull a value such as the current temperature, and write, to set one, such as a target temperature. Devices announce themselves in a standard format, so a model no longer needs a bespoke translation layer written for each machine. The spec also reserves space for natural-language tags, so a device's characteristics, safety limits and capability boundaries can be written out in terms a model can read directly. Three entry points sit on top: MCP, a command line, and a code-file API — Anthropic says this lets you orchestrate multiple devices in a single line of code.
The numbers partners are getting
Genentech used it to run BCA protein quantification, with Claude coordinating a liquid handler, a robotic arm and a plate reader at once, and working out transfer flow rates for different liquids on its own — about 140 µL/s for water (RMSE 0.016), dropping to 10 µL/s for the more viscous BSA solution (RMSE 0.181). When a tip pickup failed or liquid-level detection went wrong, the agent backed off and retried on its own.
Carnegie Mellon ran a gradient-dilution dose-response experiment, executing 3x faster, with onboarding taking just 8 hours while juggling three mutually incompatible computer interfaces at once. One detail stands out: when a fitted curve's R² fell short of 0.98, the agent tuned its own parameters and reran the fit.
QuEra's numbers are the most striking. When a quantum computer's laser lock dropped, recovery used to take an average of 150 seconds with a 58% success rate; with an AI controller, that became 6 seconds and 96%, later reaching 99.3% with no human intervention. Along the way it also retuned the PID parameters, cutting residual noise from 15.7 mV to 1.55 mV — roughly an order of magnitude. Separately, the University of Washington's Baker and Pinglay labs connected six instruments within a week, with qPCR curves monitored in real time and runs auto-stopping at the optimal amplification point; HHMI Janelia used it to chain two-photon microscope software across seven different vendors.
The value sits in the driver layer
MHS is essentially MCP's idea — giving a model one unified tool interface — moved from software into hardware. A rough estimate: a mid-sized wet lab typically runs twenty to thirty instruments; at a vendor's own pace of two weeks per device, just wiring up the interfaces would eat more than a year of labor. At the 8-hour pace, one person could clear the whole lab in a month. That's a rough estimate only — the real figure depends on whether a given device even has a programmable interface, which is exactly MHS's current hard limit.
The list of announced supporters already spans a wide range: AWS's Strands Robots, Automata's LINQ platform, Danaher, Doosan Robotics, QIAGEN, Tecan's Fluent liquid handlers, Universal Robots, plus Hugging Face's LeRobot library and Raspberry Pi. Whether a software standard actually sticks always comes down to whether hardware vendors are willing to adapt to it themselves — and this list suggests that bar is already half-cleared.
The limits are spelled out more plainly than the hype
Anthropic's own list of limitations is fairly candid: models still lack physical intuition, so questions like how bubbles form still need an expert's hint; only devices with a programmable interface can be connected at all; and having an agent watch an instrument continuously comes with a compute overhead that hasn't been fully worked out yet. On safety, each device's safety limits are enforced by the MHS driver itself, safety evaluation is being carried out jointly with research partners, and a physical-safety roadmap has to be locked down before the standard goes open source.
The standard will eventually be open-sourced, contingent on safety evaluation wrapping up, best practices taking shape, and the research preview's conclusions being published externally. For now it's application-only, with an official waitlist. For people running wet labs, the real payoff is in the iteration cycle — for the first time, an experiment might be timed from design to working run in hours rather than weeks.
Sources: Anthropic official announcement, CocoLoop, modelhardwarestandard.com; partner experimental data, the device-vendor list and the open-source timeline are all verified against the official announcement's wording.