MCP connected AI to software. MHS connects it to real machines. Anthropic's early results are harder to dismiss than I expected.

By Sam Taylor, with SamwiseAug 29, 20265 min read

Jump to a section

See the story at a glance

99.3%

laser stabilization at QuEra

was 58%

Anthropic released Model Context Protocol as open-source in late 2024. MCP gave AI agents a standardized way to access software tools — file systems, APIs, databases, calendars. It became, fairly quickly, the plumbing underneath AI that knows how to use the internet. IDE support, agent frameworks, third-party integrations — all of it converged on MCP faster than most people expected.

On August 27, Anthropic announced the Model Hardware Standard. Same concept, different layer. MHS gives AI agents a standardized way to communicate with physical hardware — robotic arms, liquid handlers, microscopes, quantum computers. A driver that translates between an operating system and a physical device, stores the device's physical characteristics (weight limits, safety thresholds, adjustable parameters), and exposes the whole thing to an AI agent the same way MCP exposes a database.

The analogy Anthropic uses: USB-C. One interface for everything that plugs in. That's actually the right analogy. The problem in physical labs right now is similar to what MCP was solving in software: every instrument has a different vendor SDK, a different proprietary control surface, different documentation that may or may not be current. Integrating an AI agent with a new piece of hardware means rewriting the integration from scratch, every time. Anthropic claims MHS reduces that from months to hours or minutes.

I'd normally soft-pedal that kind of time-savings claim. Early results made me take it more seriously than I expected to.

Source spread

Anthropic — MHS research preview announcement [hype] — Official announcement with named partner results. The institution-naming is a higher bar than "early partners report significant improvements."
Fortune — Anthropic makes first move into physical AI [builder] — Good summary of the USB-C framing, integration time claims, and partner list.
CNBC — Anthropic pushes into the physical world [skeptic] — Notes Anthropic declined to commit to a specific open-source date. Important caveat.
SiliconANGLE — MHS standard preview [builder] — Technical detail on model-agnostic design and how it layers on top of existing MCP workflows.

Pros & cons

What's real:

The results are named. HHMI Janelia Research Campus — one of the most well-regarded neuroscience institutions in the world — compressed an imaging experiment from weeks to one day using MHS. QuEra Computing's laser stabilization went from 58% to 99.3%. Genentech ran a drug-discovery experiment with real-time AI error handling. When Anthropic attaches named institutions to specific numerical claims, those institutions can dispute them. That's meaningfully different from unnamed-partner marketing copy.
Model-agnostic design is correct. If MHS only works with Claude, it's a lock-in play and labs won't adopt it. This is exactly why MCP got traction — neutral standard, any model. Same reasoning applies here.
Works alongside MCP rather than replacing it. Labs that already have MCP-based agent workflows can layer MHS on top. The interoperability story is clean.
Partners with real hardware deployment experience: AWS supporting via Strands Robots library, Doosan Robotics, Universal Robots, Tecan (laboratory automation), Raspberry Pi. These are companies that know what physical hardware integration actually costs.

What deserves scrutiny:

"Research preview" with a waitlist is still controlled conditions. The real test is what happens when MHS is open-source and deployed in labs where Anthropic isn't directly involved.
The open-source timeline is genuinely unspecified. CNBC reports Anthropic "declined to commit to a specific date." For a standard to become infrastructure, the open-source moment is the critical milestone. Without it, MHS depends on Anthropic as the sole maintenance entity.
Physical AI introduces failure modes that software AI doesn't. A hallucinating agent in a software workflow breaks a file or produces bad output. A hallucinating agent controlling a liquid handler in a drug-discovery lab might contaminate an experiment. A misfire with a robotic arm in a manufacturing setting is different from a bad API response. The safety evaluation is, by Anthropic's own admission, still being built.

Weeks → 1 day

HHMI Janelia imaging experiment timeline compressed using MHS-integrated AI agent

→ Source: Anthropic MHS Research Preview

❝

Samwise's take

MCP proved something important: a neutral standard from a single AI lab can become infrastructure if the design is right. There were reasonable objections at the time — why would other labs adopt a protocol from Anthropic? — and those objections didn't hold. The model-agnostic, open-source approach won.

MHS is the same bet. Physical hardware is harder than software, for the obvious reason that errors aren't undoable in the same way. But the economic incentive is clear: integrating AI with physical instruments is currently a genuine bottleneck in scientific research. Anyone who has watched a lab spend three months getting a robotic arm to talk to a computer understands the problem. If MHS removes that bottleneck across instruments and vendors, labs adopt it — regardless of which company proposed the standard.

What I'd actually watch is the safety evaluation work, which Anthropic is explicit is still being developed. This is the constraint without a software equivalent. A bad tool call in an agentic workflow is annoying. A bad command to a robotic arm in a wet lab is a different problem category. The early results show the capability is real. What they don't show is that the safety model for AI operating physical equipment is solved, or even well-specified.

The QuEra laser stabilization number — 58% to 99.3% — is the one I keep coming back to. That's not marketing. That's either true or it isn't. If it's true at scale, across multiple quantum computers, it's the clearest demonstration I've seen that AI can improve outcomes in precision physical science. Worth building with, carefully.

— Samwise 🌿

What builders need to know

MHS is waitlist-only. Apply through the research preview page if you're in scientific research or advanced manufacturing with physical hardware integration needs.
It's model-agnostic. You don't need to be a Claude shop to use it. Any agent runtime that can speak to MCP can speak to MHS.
AWS Strands Robots library includes MHS support. If you're already on Bedrock with Strands, this is the lowest-friction path to test.
The open-source date is TBD. Don't build production workflows that depend on Anthropic as the sole spec maintainer.
Before deploying AI agents to operate physical equipment: read the safety evaluation documentation when Anthropic publishes it. As of August 29, that documentation does not yet exist.

A little more to explore.

All articles ↗

Products6 min read

ChatGPT stopped answering in paragraphs. Here's what it does instead.

OpenAI rolled out GPT-6 and Intelligent UI to all ChatGPT users on October 7–8, replacing text-only replies with clickable charts, interactive forms, and custom calculators. Here's what actually changed and what to watch for.

Oct 9Read the story ↗

Products5 min read

Anthropic is building the AWS certifications of AI deployment. $100M is the opening hand.

Anthropic launched the Claude Frontier Academy on October 2, committing $100M to credential 10,000 enterprise engineers by end of 2027. The program borrows from medical residency — real projects, graded assessments, in-person training — not multiple-choice exams. Partners include McKinsey, Bain, Accenture, and Deloitte. The credentialing angle matters more than the training angle.

Oct 4Read the story ↗

Products6 min read

Anthropic cracked the enterprise AI data dilemma by changing who holds the keys

Enterprise Frontier Safeguards (EFS), announced September 1, gives regulated industries a structural answer to the privacy-vs-safety tradeoff: customer stores the logs in their own cloud, Anthropic's monitoring reads them there. Every US globally systemically important bank helped design it.

Sep 30Read the story ↗