Anthropic's Model Hardware Standard: An MCP for the Physical World
Anthropic has announced the Model Hardware Standard (MHS) — a shared specification that lets AI agents safely operate physical devices, with the goal of compressing lab integration work from "weeks to months" down to "hours to minutes." The announcement (August 27) came with measured results from six pilot institutions, spanning protein quantification to quantum laser re-locking.
What MHS Actually Is
MHS is a pure software specification — it defines no voltages, connectors, or mechanical dimensions. It only solves one problem: how agents and equipment talk to each other. It has four layers:
| Module | What it does | Example | |---|---|---| | Standard drivers | Translate hardware into two primitives | Read temperature, set temperature | | Device discovery | Devices self-describe in a standard format | Agents find devices on the network automatically | | Metadata tags | Natural-language hardware specs, auto-generated reference files | Payload limits, adjustable parameters, safety limits | | Three control channels | MCP, command line, direct API calls | Orchestrate a chain of devices in one line of code |
Safety limits are mandatory and enforced at the driver layer. The stack is model-agnostic — Claude can use it, but so can agents from other vendors.
The origin is notably grassroots: Janelia postdoc Arco Bast built a shared-memory dictionary for his own brain-imaging rig; Anthropic's Alek Kemeny saw it, and the two grew it into a standard.
Six Pilots, Hard Numbers
Pilot list: Genentech, University of Washington, Carnegie Mellon, HHMI Janelia, QuEra, Tetsuwan Scientific.
Carnegie Mellon — Serial Dilution
- CyBio Felix liquid handler, Varioskan LUX plate reader (no API — the agent drove its GUI directly), Spinnaker robotic arm, cameras, across three incompatible computers — all wired in ~8 hours (previously billed by the week).
- Researchers planted six fault types (missing plate, misaligned plate, occupied reader, camera disconnect, device loss, e-stop) — all caught.
- When one run's R² fell below 0.9, the agent judged it a failure, changed concentrations itself, and re-ran to R² > 0.98, with no human intervention.
- Legacy script: 150 seconds to recover, 58% success rate; human experts: 5–10 minutes per event.
- The agent-written re-locking controller: 695/700 successful trials, recovery in seconds.
- 363 experiments over 16 hours unattended; residual lock error reduced from 15.7 mV to 1.55 mV. Over 19 hours it never lost tuning; a human expert loses it ~1.6 times per hour.
- Claude orchestrated three devices and optimized its own pipetting speeds: 140 µL/s for aqueous solutions, 10 µL/s for viscous BSA — matching experienced technicians.
- Self-recovered from tip-pickup and liquid-level detection errors. The one moment requiring humans: it couldn't tell whether a bubble problem was physical or software — a person had to clarify.
- Claude's spatial and physical reasoning still needs expert supervision — QuEra's agent was even over-cautious, stopping at night to wait for approval;
- Long-running agents carry nontrivial compute costs;
- Orchestration of complex protocols still requires tuning.
- Anthropic announcement: https://www.anthropic.com/news/model-hardware-standard-research-preview
- Application portal: https://modelhardwarestandard.com
QuEra — Laser Re-locking for a Neutral-Atom Quantum Computer
Genentech — BCA Protein Quantification
Ecosystem
Vendors engaging: AWS, Danaher, Tecan, QIAGEN, Doosan, Universal Robots, Automata, MBF Bioscience.
Early adopters: Hugging Face (via its LeRobot robotics library) and Raspberry Pi, which shipped an official MHS driver for its camera.
Why It's Bigger Than Lab Automation
While embodied AI's main narrative is giving models bodies (humanoids, end-to-end VLA, world models), MHS takes a different path: standardize the *agent ↔ physical device* interface first — closer to what USB did for peripherals and MCP did for tool calling. Route one is sexy; route one is expensive. Route two targets the installed base: billions of programmable devices already sitting in labs and factories that lack not intelligence, but a unified thin interface.
The research impact is direct: at the University of Washington, a de novo protein design costs one cent to generate, but ~$100 and a week to validate. The bottleneck is entirely wet-lab side. The QuEra case even suggests that quantum computer operations could be one of AI agents' first real jobs.
The Cold Water
Anthropic itself acknowledges three shortcomings:
One more from the original author: MHS currently requires devices with programmable interfaces (GUI counts), leaving purely mechanical devices unsolved. Anthropic says it's working with vendors whose equipment lacks interfaces; progress is unclear.
The Takeaway
The most important detail is the open-source pledge. Closed standards don't generate network effects; open ones can become de facto standards. Anthropic's ambition isn't selling hardware — it's defining the device protocol layer for the agent era: the MCP of the physical world. The signal to watch over the next six months: whether "MHS compatible" becomes a printed selling point in Tecan and Danaher product catalogs.