Anthropic Built MCP for Centrifuges. You Can't Have It Yet.

TL;DR

On August 27 Anthropic kicked off the first phase of a research preview for the Model Hardware Standard (MHS), a shared specification that lets an AI agent operate microscopes, centrifuges, pipette robots, spectrometers, incubators, and robot arms through one interface. It is built on top of MCP, it is model-agnostic, and it started as a collaboration with HHMI Janelia Research Campus. The headline number comes from QuEra Computing: a laser stabilization routine that a hand-written script landed 58% of the time now lands 99.3% of the time. The catch is at the bottom of the page. You cannot download any of it.

The driver is the whole idea

Every lab instrument ships its own world: a vendor SDK, a serial protocol, a Windows-only GUI, and if you are unlucky, a dongle. Anthropic says setting up and integrating that hardware normally takes a facility weeks or months, and MHS is meant to compress it to hours or minutes.

It does that by putting a driver in front of each device that exposes a deliberately tiny primitive set: read (for example, get temperature) and write (for example, set temperature). Everything else an agent needs to know lives in natural-language tags inside the same driver, so "this incubator's latch sticks below 4C" becomes a documented device property instead of tribal knowledge held by the one postdoc who knows. You can write those tags yourself, or, per Anthropic, hand the job to an agent that interviews you about your setup.

If you have ever installed a printer, you already know the shape of this. Twenty years ago every manufacturer spoke its own control language, then a driver layer turned all of it into one verb: print. MHS is the same trick, except the verbs are read and write, and the printer is a confocal microscope that costs more than a house.

Agents reach that driver three ways: an MCP server, a command line interface, or code files that expose it as a plain API. That is a sane spread. Not every workload wants a model in the loop on every step, and a calibration sweep is much better off as a script that the model wrote once.

58% to 99.3%, on a quantum computer's lasers

The most concrete result in the announcement comes from QuEra, which builds neutral-atom quantum computers. Locking a laser to a reference is a fiddly feedback problem, and QuEra's existing approach was a bespoke script. Anthropic reports it succeeded 58% of the time and took roughly 150 seconds per attempt. Claude, driving the same rig through MHS, reports a 99.3% success rate in 0.9 to 14 seconds depending on how badly the loop had been knocked out of alignment.

The same writeup reports Claude retuning the loop's PID parameters and cutting residual error from 15.7 mV to 1.55 mV, roughly a tenfold improvement. This is the least glamorous and most useful category of AI work: a search over a parameter space that a human would sample maybe thirty times before deciding the current settings are fine, actually.

Read these as vendor-reported figures. They come from Anthropic's own writeup of partner deployments, not from independent replication, and nobody outside the preview can run the setup to check.

Eight hours instead of several weeks

The integration numbers are the part that would change a lab's week. Anthropic says Carnegie Mellon stood up an MHS integration in about eight hours where a vendor-built setup would have taken several weeks, and then ran dose-response experiments roughly three times faster than its previous method. The University of Washington Baker and Pinglay labs connected six instruments in under a week.

At Genentech, the agent optimized liquid handling for a BCA protein assay and landed on about 140 microlitres per second for water and about 10 microlitres per second for a viscous BSA solution. The point is not either number. It is that a single hardcoded flow rate is exactly the kind of thing a rushed protocol keeps forever, and the agent went and found that viscosity changes the answer by more than a factor of ten.

Tetsuwan Scientific ran the largest-sounding pass: 9,143 individual dispenses across 300 unique transfer types and 1,508 measured conditions, producing precision predictions Anthropic says beat manufacturer specifications by roughly 12% on 31 of 45 runs, and around 17% on the most-replicated data.

Model-agnostic, vendor-agnostic, and not yours

Two design choices matter more than any benchmark here. MHS is model-agnostic: Fortune reports it works with any LLM, including OpenAI's models and open-weight ones, not just Claude. And Anthropic's stated goal, per its staff, is to avoid vendor lock-in for scientists, with an eventual open-sourcing of the standard so any device maker can adopt it. Anthropic's Alek Kemeny described MCP to CNBC as "kind of like the USB for AI to software connection", which tells you exactly which analogy the company is aiming for.

The ecosystem list is doing a lot of the persuading: Amazon Web Services, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots, Hugging Face, and Raspberry Pi. That spans the pipetting-robot incumbents, an industrial arm vendor, and the board in your drawer, which is roughly the coalition a hardware standard needs if it wants to survive contact with procurement.

But "eventually" is carrying weight. Today MHS is an invite-only research preview with an application form, no public specification, no repository, and no announced license. Anthropic frames the phase as time to build safety evaluations and guidelines for AI systems controlling physical equipment before a wider release, which is a defensible reason to keep the door shut. It is still a closed door.

Why this is different from an agent clicking buttons

Agents have been driving software for two years. The interesting change is not capability, it is the shape of the failure surface. An agent that hallucinates a citation costs you an embarrassing footnote. An agent that hallucinates a centrifuge speed costs you a rotor, and possibly the wall behind it.

That is why the natural-language tags are the sleeper feature. A traditional driver encodes what a device can do; MHS asks you to also encode what it must not do, in the same file, in a form the model actually reads. Whether that holds up under an agent that is confidently wrong is precisely the thing a safety preview should be measuring, and precisely the thing none of us can measure from outside.

If you build agent infrastructure, the practical move is to watch two things: whether the spec lands under a real open license, and whether the primitive set survives contact with devices that do not fit read and write. Plenty of instruments have stateful, long-running, interruptible operations that a two-verb interface handles badly. The standard either grows a third primitive or grows a pile of conventions, and which one it picks will tell you how seriously to take it.

Key Takeaways

• MHS is a driver layer, not a model. Devices expose read and write primitives plus natural-language tags, and agents reach them through MCP, a CLI, or code files.
• The best result is a control-loop tune. QuEra's laser lock went from 58% to 99.3% success and from about 150 seconds to 0.9 to 14 seconds per attempt, per Anthropic.
• Integration time is the real pitch. About eight hours at Carnegie Mellon versus several weeks for a vendor-built setup, and six instruments in under a week at the University of Washington.
• Model-agnostic by design. It works with non-Claude models, and Anthropic says it plans to open-source the standard so any device maker can adopt it.
• Nothing is public yet. Invite-only preview, no spec, no repo, no license, and every number comes from Anthropic's own writeup of partner deployments.
• Watch the primitive set. Long-running, interruptible instrument operations do not map cleanly to two verbs, and how MHS handles them will decide whether it generalizes.

Sources: Anthropic, Previewing the Model Hardware Standard, modelhardwarestandard.com, CNBC, Fortune