Claude can already open files, call APIs and operate software through tools such as MCP.

Physical machines are harder.

A microscope, robotic arm and plate reader may come from three different companies. Each has its own software, commands and data formats. Making them work together can require weeks of custom integration.

Anthropic opened a research preview of the Model Hardware Standard.

MHS gives AI agents a common way to discover and operate programmable machines. Anthropic is testing it with scientific labs, robotics companies and advanced manufacturers.

This is not Claude controlling every robot.

It is an attempt to give models a standard interface to the physical world.

I am Alex. Welcome to ShortCu8 by Innov8.

Lets Dive Deep🐰🐰.

Think of physical AI as a four-part loop:

Model -> MHS -> Machine -> Measurement -> Model

The model decides what should happen.

MHS translates that decision into allowed hardware commands.

The machine performs the action and sends its state or measurements back.

The model reads the result, then continues, adjusts or asks a human for help.

Without the middle layer, every model and every machine needs a custom integration.

Software already has APIs.

Physical equipment also has programming interfaces, but there is no shared language across manufacturers.

Anthropic says MHS solves this through standard drivers.

Each driver describes the machine's states and procedures.

A state could be:

A procedure could be:

The driver can also contain physical information that ordinary code may not reveal, such as the weight of a robotic arm or the safe temperature range of an instrument.

Claude can access those devices through MCP, a command line or code. MHS is model-agnostic, so the standard is not limited to Claude.

Genentech tested MHS on a workstation containing three machines: a liquid handler, a robotic arm and a plate reader.

The experiment measured protein concentration in a sample.

A scientist described the protocol to Claude. Claude planned the run, instructed the liquid handler, moved the plate using the robotic arm and collected measurements from the reader.

Part of the process ran as a closed loop.

Claude selected a liquid-transfer speed, performed the transfer, measured the result against an expert example and adjusted the speed.

It also failed in a very physical way.

Claude initially used similar settings for water and a thicker protein solution. The protein solution formed bubbles, damaging the accuracy of the transfer.

Claude treated the resulting errors like a software problem and kept retrying. Human researchers had to explain that the bubbles were a physical failure and that the liquid needed gentler handling.

Once that lesson was provided, the team converted it into a reusable skill for later experiments.

That small failure tells us what MHS can and cannot do.

Claude can coordinate the machines and improve from measurements. It does not automatically understand all the physics happening inside them.

QuEra Computing used MHS to give an AI agent access to part of the laser system inside its quantum computers.

The lasers must maintain an extremely precise frequency. When the system loses that frequency lock, it needs to recover before the machine can continue operating.

QuEra already had a recovery script built by specialists. According to the company, that script recovered the laser 58% of the time and took roughly 150 seconds per attempt.

During the MHS pilot, Claude repeatedly disturbed and recovered the system overnight. It converted the original linear procedure into a decision tree that reacted differently depending on what the instruments reported.

In the development run, recovery improved to approximately six seconds with a 96% success rate. Anthropic reports that a later blind test reached 99.3%.

These are results from an early partner pilot, not an independent benchmark.

But they show the important mechanism: the agent could act, measure, compare and turn what it learned into deterministic code.

MHS could become useful even when Claude is not the model operating the machine.

The specification is designed to be model-agnostic. A hardware company could build one MHS driver instead of creating a separate integration for every AI agent.

Hugging Face is adding MHS support to its LeRobot library. Raspberry Pi is working on integrations for some of its products. Universal Robots, Doosan Robotics, AWS and several scientific-equipment companies are also participating in the preview.

If that adoption continues, an agent could discover a new machine, read its capabilities and operate it without engineers rebuilding the entire connection each time.

That is the same reason MCP became important for software agents.

A good model is useful.

A shared protocol lets that model reach an ecosystem.

MHS has not been open-sourced yet.

It currently requires hardware with a programmable interface. It cannot make a machine controllable when that machine exposes no software connection.

Anthropic also says Claude's spatial and physical reasoning still requires expert oversight. A wrong software action may corrupt a file. A wrong hardware action can damage equipment, waste a biological sample or injure someone.

The hard part is therefore larger than connecting the machine.

The standard must also describe permissions, operating limits, human approvals and what the agent should do when the physical world behaves differently from its internal model.

MCP gave AI agents a common way to use software.

MHS is trying to do the same for physical machines.

That changes the consequence of an agent action.

An AI working inside a computer moves information. An AI connected through MHS can move a robot arm, change a laser or run an experiment.

The standard connecting models to machines may eventually matter as much as the model itself.

Because the next important AI agent may not be sitting inside a chatbox.

It may be working through the night inside a laboratory.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Try it for FREE today

#AD2

Most SaaS finance teams know their billing process is slow.Most SaaS finance teams know their billing process is slow. Few know what it's costing them.

The Tabs Billing Lag Calculator puts a dollar figure on it in two minutes — benchmarked against top SaaS companies.

See your number.