Most of the built environment runs on machines that cannot talk to software. Total stations, batching plants, monitoring sensors and lab rigs for materials testing each arrive with their own vendor programme, data format and quirks. Getting any two of them to cooperate often means paying a specialist to write glue code. That integration tax is one big reason AI on-site has stayed stuck at the reporting layer.
On 27 August, Anthropic opened a research preview of something aimed squarely at that problem: the Model Hardware Standard (MHS), a shared specification that lets AI agents discover, understand and safely operate physical devices. It began as a collaboration with HHMI Janelia Research Campus that works with any device that has a programmable interface. It’s also model-agnostic, which means any agent can reach it through standard protocols such as the Model Context Protocol.
If MCP was the USB port for AI and software, MHS is the attempt to build the same port for hardware. Anthropic says integration work that takes labs weeks or months today drops to hours or minutes. For an industry that owns more machinery per pound of revenue than almost any other, that claim deserves attention.
A driver, not a robot
MHS is plumbing rather than robotics. It introduces a standardised driver built on simple primitives – read a value, write a value – that any device can act on, and it makes each device discoverable in a common format so agents and machines can find each other across a network without a bespoke translator in between.
The clever part is how the driver captures the knowledge that normally lives in paper manuals and someone's head. Users describe their equipment in plain language (or let an agent interview them about it), and the driver turns that into a reference file covering what the device can measure, what can be adjusted and what safety limits will be enforced. Anthropic gives the example of a robot arm's weight – information an agent cannot infer from code, and exactly the sort of thing a site engineer knows about a piece of plant that its API never will.
What the early adopters found
Anthropic shared MHS with labs and manufacturers ahead of the preview, and the partner write-ups on its announcement carry real numbers. All of these are partner-reported figures published via Anthropic, so read them as promising pilots rather than audited benchmarks.
Researchers at Carnegie Mellon University wrote drivers from scratch for a liquid handler, a plate reader, a robotic arm and monitoring cameras, plus an orchestration layer, in about eight hours – against the several weeks a vendor-built setup typically takes.
QuEra Computing had Claude develop a controller for re-locking the lasers inside its quantum computers, and reports the finished script recovered the lock in 695 of 700 validation trials, taking seconds where a human expert needs 5 to 10 minutes.
Genentech used MHS to have Claude orchestrate a standard protein assay across three instruments, with the model autonomously optimising liquid-handling parameters that would normally need a specialist writing custom logic.
Zihao Song, a PhD student in the University of Washington's Baker and Pinglay labs, connected six instruments in under a week, including the time spent writing drivers.
Song's summary is the clearest statement of what the standard actually does: "MHS essentially gave the agent eyes, hands, and a sense of timing: it could see the status of every instrument, run each one, and coordinate them to work together."
With MHS, I now look forward to the day when hardware control no longer limits the questions I can ask.
Why construction should care about lab kit
None of the launch partners pours concrete. Look at who is building support in, with Tecan and QIAGEN adding it to their instruments and AWS doing so through its Strands Robots library. Universal Robots and Doosan Robotics are adding support for robotic arms, with Hugging Face doing the same in its LeRobot robotics library. Raspberry Pi is also adding support across a number of its products, with Raspberry Pi boards and industrial sensors already everywhere on sites and in structural monitoring. Once device makers ship a common agent interface as standard, the gap between an AI that summarises the monitoring report and an AI that reads the sensor, adjusts the parameter and flags the anomaly at 3am gets very short indeed.
The plant side is moving on its own timetable. Caterpillar, as we cover elsewhere this week, is rolling its decades of mine automation into construction, with its Cat AI Assistant now in the hands of technicians and 1.6 million connected assets feeding the data.
The hard part about autonomy and about physical AI is incorporating that technology into the customer jobsite and into the workflows.
A shared interface standard is precisely the kind of thing that makes that hard part cheaper.
The awkward timing on safety
An honest read has to note the calendar. Days before this announcement, Anthropic was publishing a post-mortem on three July incidents in which Claude models gained unauthorised access to real computer systems during evaluations, alongside a separate incident reported by the UK AI Security Institute. Handing agents control of lasers and robot arms in that same season will raise eyebrows, and reasonably so.
To its credit, Anthropic is gating the open-source release on exactly this: the preview exists to build safety evaluations and best practices with partners first, with safety limits enforced at the driver level, independent of the model. The partner accounts are candid about limitations too – Genentech had to teach Claude that foaming samples were a physics problem rather than a software bug, and QuEra kept a laser engineer reviewing every step. Expert oversight is doing real work in all of these pilots.
Takeaway
Standards are boring until they win, and then they are everything. MHS is a research preview, its numbers are partner-reported, and construction equipment is a harder, dirtier, more safety-critical environment than any lab bench.
The direction is unmistakable. The interface layer between AI agents and physical machines is being built now, with serious hardware names already on board.
The open-source opportunity is worth watching. When MHS goes open source, the firms that understand their own kit well enough to describe it to an agent will move first.
Know your equipment. It is worth knowing which of your machines have a programmable interface. Someone will ask soon.
The line between AI that writes about the work and AI that touches the work is getting thinner every week, and we track it so you don't have to. For the stories where software meets steel, sensors and site kit, join thousands of built environment professionals reading Project Flux.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
