Anthropic's Model Hardware Standard (MHS): The New Standard for AI Agents That Operate Real Machines

Published August 27, 2026By ABD Legacy LLC
Model Hardware Standard MHS Anthropic physical AI AI agent standards AI agents for manufacturing

On August 27, 2026, Anthropic opened a research preview of the Model Hardware Standard (MHS) — a shared specification for AI agents to safely operate physical devices, starting with scientific research labs and advanced manufacturers. Early partners report results: Genentech ran an assay with an agent recovering from its own hardware errors, HHMI Janelia compressed a half-day microscope setup into a single step, and QuEra lifted quantum laser lock recovery from 58% to 99.3%.

This matters beyond the lab. MHS is Anthropic's agent-standards playbook extended from software to the physical world — and agencies advising research, manufacturing, and robotics clients should understand it.

What is the Model Hardware Standard (MHS)?

MHS is a standardized driver layer: software that translates between a computer's operating system and a hardware device, using "a simple set of primitives—commands like 'read' (for example, 'get temperature') or 'write' (for example, 'set temperature')—that any hardware device can understand and act on."

A driver standard, not a robot

MHS ships no robots or instruments — it makes existing devices speak one language: devices are "discoverable in a standard format, so that devices and agents can find each other and communicate across networks without needing a bespoke 'translator' program in between." Devices carry natural-language tags — like a robot arm's weight or its safety limits — that generate a reference file describing the device's general characteristics, so an agent can operate hardware it has never seen before.

Model-agnostic and built on MCP

MHS "works with any device that has a programmable interface," is model-agnostic, and "any agent harness can access it using standard protocols, such as the Model Context Protocol." Anthropic open-sourced MCP in 2024 to connect agents to software and data; MHS is the same idea pointed at machines.

Why Anthropic is moving into physical AI

Fortune called MHS "Anthropic's first foray into so-called physical AI." CNBC framed it as the company pushing "into the physical world" — and noted the stakes: rivals OpenAI and Amazon "have spent billions of dollars designing AI-native devices and manufacturing tools," while Anthropic is reportedly building a silicon team and has hired hardware executive Caitlin Kalinowski, formerly of OpenAI, Meta, and Apple (CNBC).

Elizabeth Kelly, head of beneficial deployments, told CNBC: "We built this for science to sort of show the promise of AI, but there's also huge benefits here for enterprise and for industry." Co-creator Alek Kemeny told Ars Technica: "This idea could be used to have AI run any science experiment in the world."

The roadmap mirrors MCP's 2024 playbook — research preview now, open source later: Anthropic is sharing an early version with partners "ahead of making the standard open source," and CNBC reports it "eventually plans to open-source the standard, which means any device manufacturer in any industry will be able to adopt it."

How MHS works in practice

Agents drive MHS devices through three control mechanisms — the Model Context Protocol, the command line, and code files (APIs) — with multi-device orchestration from a single line of code.

The integration claim is the business hook: setups that typically take "weeks, if not months" are reduced "to hours or minutes." The payoff: "autonomous, round-the-clock experiments and workflows, with agents able to reason through each step in an experiment, update parameters in real time, and, in some cases, recover from hardware errors without intervention." Self-recovery is what separates MHS pilots from ordinary automation.

Early results from the research preview

Genentech: an autonomous BCA assay

Genentech automated the BCA protein assay — a standard total-protein-measurement procedure coordinating a liquid handler, a robotic arm, and a plate reader. Claude orchestrated the three instruments and converged on optimal flow rates: water at roughly 140 µL/s (0.016 RMSE) and viscous BSA at roughly 10 µL/s (0.181 RMSE). When the agent "encountered several unexpected errors, including tip pickup failures and fluid detection errors," it "managed to recover on its own—a capability that current scientific instruments mostly lack."

The caveat: Claude initially misread bubble and foam errors as software bugs until Genentech corrected it — hence Anthropic's "we have more work to do on the standard before we open-source it."

HHMI Janelia: microscopy rigs that talk to one agent

MHS began as a collaboration between Anthropic's Alek Kemeny and Arco Bast, a postdoctoral scientist at Virginia's HHMI Janelia Research Campus. Researcher Virginie Ruetten used MHS "to unify and orchestrate a rig that previously involved seven different vendor programs without a shared interface." Bast's custom microscope was the first rig to run on it, with Claude aligning the beams, tuning the optics, and checking its own results against the sensors — "turning a half-day of manual setup into a single step."

QuEra: laser lock recovery, 58% to 99.3%

QuEra builds neutral-atom quantum computers whose lasers must hold an ultra-precise frequency "lock." A bespoke recovery script "only worked about 58% of the time, and it took around 150 seconds per attempt." After an overnight agent loop with MHS, recovery took about six seconds and worked 96% of the time in the development run; in a later blind test across 700 trials, the agent "recovered the correct lock 695 times, a 99.3% success rate." QuEra's blog notes the controller is "a deterministic, fully inspectable program."

University of Washington and Tetsuwan Scientific

Two smaller pilots round out the preview: University of Washington PhD student Zihao Song used MHS for a remote instrument dashboard, an agent-supervised qPCR run, and collision-free plate handoffs; Tetsuwan Scientific ran a citizen-science qPCR workflow on its ResearchOS platform in San Pedro Creek, California.

The vendor ecosystem already signing on

Anthropic's announcement lists AWS (Strands Robots), Automata, MBF Bioscience, QIAGEN, Tecan, Universal Robots (adding support to its robotics platform), Hugging Face (LeRobot), and Raspberry Pi (Camera MHS Driver); Fortune's list adds Carnegie Mellon, Doosan Robotics, and Danaher. Anthropic's Jonah Cool on the motivation: "We want to avoid vendor lock-in for scientists." Fortune also notes Nvidia has long championed physical AI, with Jensen Huang predicting "every industrial company will become a robotics company."

MHS vs MCP: two standards, one agent story

Kemeny's shorthand: MCP is "kind of like the USB for AI to software connection." MHS extends that idea to physical equipment, and the two stack — agents reach MHS devices through MCP itself. The agent-standards story is no longer software-only — see our AI agent standards guide for the full map of MCP, Agent Plugins, AGENTS.md, and MHS.

What MHS means for AI agencies

A new client vertical opens up

Physical AI is now addressable by small integrators, not just equipment OEMs. If a client's instruments have programmable interfaces, an agent can orchestrate them — the kind of build an agency can scope, pitched against the "hours or minutes" integration claim.

Procurement decisions lock in agent compatibility

Standards adoption is a buying signal: as vendors add MHS drivers, instrument purchases become agent-compatibility decisions. Track who ships MHS support, and advise clients to demand documented, safe interfaces in every RFQ.

Oversight is still the product

Anthropic is explicit that expert oversight is still required — Genentech's bubble episode shows the failure mode. For agencies selling agent adoption, the durable value is the governance layer: safety limits declared in device tags, human-in-the-loop checkpoints, and audit trails. That's where margins live once standards commoditize the plumbing.

The open-source and safety roadmap

MHS is a research preview, not a shipped product — every result above is a partner pilot. Anthropic is using the preview to "collaborate to build safety evaluations and develop best practices for AI systems operating physical equipment," and is "developing a physical safety roadmap" tied to its safeguards policy. WIRED flags the stakes: letting AI use physical systems "raises the prospect of new risks because of the potential to damage physical systems or hurt people."

Helping a lab or manufacturer evaluate AI agent integrations? An AI automation agency can scope the hardware, the oversight layer, and the build.

Find an AI Agency →

Or read why agent standards are becoming an agency service line.

Frequently asked questions

What is Anthropic's Model Hardware Standard (MHS)?

A shared specification Anthropic announced August 27, 2026 for AI agents to safely operate physical devices such as microscopes, liquid handlers, and robotic arms. In research preview now; open source later.

How does MHS work?

A standardized driver layer that translates between an agent and a hardware device using simple primitives like read ("get temperature") and write ("set temperature"). Devices carry natural-language tags for weight and safety limits; agents control them via MCP, the command line, or code files.

Is MHS open source?

Not yet. Anthropic is sharing an early version with partners to build safety evaluations and best practices before open-sourcing — the same path as MCP in 2024.

What is the difference between MHS and MCP?

MCP connects AI agents to software and data sources; MHS extends the same idea to physical hardware. MHS is model-agnostic and accessible through MCP, so the two work together rather than compete.

Who is using MHS?

Genentech, HHMI Janelia, QuEra, the University of Washington, and Tetsuwan Scientific, plus hardware vendors AWS, Universal Robots, Doosan Robotics, Danaher, Tecan, QIAGEN, Hugging Face, and Raspberry Pi.

Related reading

Sources

Accuracy note: MHS mechanics, partner results, and quotes are from Anthropic's August 27, 2026 research-preview announcement unless attributed otherwise. The QuEra 96% figure is the development run; 99.3% is the later blind test (700 trials, 695 successes) — Anthropic distinguishes the two, and so does this post. The silicon-team and Kalinowski details are CNBC-reported via LinkedIn. MHS is a research preview, not a shipped product; all results are partner pilots, and Anthropic states the standard still requires expert oversight.