Anthropic's Model Hardware Standard: what happens when AI agents can use physical tools?

It's early days, and access is currently limited while Anthropic works with partners on safety evaluations and best practices. But it opens up some fascinating possibilities.  

What is MHS? 

Think of it as a standard way for AI agents to understand and interact with physical equipment. 

Today, if you want software, sensors, robots, lab equipment, or manufacturing devices to work together, you're often dealing with a patchwork of custom integrations. MHS aims to provide a shared interface that makes devices discoverable and usable by AI systems.  

Anthropic says it works with any device that has a programmable interface and is model-agnostic, meaning it's not tied to a specific AI model or agent framework. Agents can access it through standard protocols, including MCP 

Is this “AI can now control everything”? 

Not really. 

I suspect we'll see plenty of dramatic headlines suggesting AI agents can suddenly control any robot or machine in the real world. That's probably not the most useful way to think about it. 

A more accurate description is that the owners of physical devices can use MHS to advertise what those devices can do, how they can be operated, and what safety constraints apply. An AI agent can then discover and use those capabilities if it's been granted access. 

In other words, MHS doesn't magically give AI control over the physical world. It provides a standard way for devices and agents to communicate. 

What could you do with it? 

This is where things get interesting. 

It’s uses extend beyond manufacturing and laboratories and can apply to everyday uses like gardening. 

Even something as simple as a Raspberry Pi connected to non-AI devices could become much more useful. Imagine a greenhouse fitted with cameras, light sensors, temperature sensors, soil moisture sensors, and maybe even an irrigation pump. 

Once those devices were discoverable through MHS, you could ask an AI: 

"Find me the watering regime that produces the best tomato growth using no more than five litres of water a week." 

Instead of simply reporting data, the agent could run experiments, adjust conditions, measure outcomes, and improve its approach over time. 

Basically, the AI could become an experimental greenhouse manager. 

Could this work outside the garden? 

Absolutely. 

With 3D printers becoming more affordable and smart home devices becoming more common, there are lots of opportunities for experimentation. 

Imagine asking: 

"Why is this room colder than the rest of the house? Run experiments this week to test heating schedules, sunlight levels, ventilation, and any other likely causes. If you need additional sensors, recommend them, but keep the investigation under a set budget." 

That's a very different type of interaction from today's smart home experiences. 

Instead of setting rules yourself, you're defining an objective and letting the system investigate possible explanations. 

Haven't we seen something similar before? 

Sort of. 

Avanade's RHS Chelsea Flower Show gold medal-winning garden in 2025 used Azure AI services, sensors attached to trees, cameras, and retrieval-augmented generation (RAG) to create a conversational experience. Visitors could effectively “talk” to trees or ask the garden how many bees were pollinating flowers. 

What made that interesting was the connection between the physical environment and the AI's understanding of it. 

It wasn't an agentic system. The AI could observe and explain, but it wasn't controlling the environment. 

If you extended that concept with an agentic framework and a standard such as MHS, you could imagine connecting grow lights, irrigation systems, and other equipment. The AI wouldn't just tell you what's happening in the garden. It could experiment with ways to improve plant health, water efficiency, pollination, or growth. 

What are the cyber security risks? 

This is probably the most important question. 

Whenever we connect AI systems to physical devices, the potential impact of mistakes or malicious activity increases. 

A compromised chatbot is one thing. A compromised robotic arm, manufacturing system, laboratory device, or building management system is something else entirely. 

That's one reason Anthropic is starting with a limited research preview and focusing heavily on safety evaluations before open sourcing the standard 

In practice, organisations will need strong identity controls, device authentication, permissions management, monitoring, and auditing. Just because an AI agent can discover a device doesn't mean it should be allowed to use it. 

The cyber security challenge isn't really the standard itself. It's making sure the right agents can perform the right actions on the right devices, with clear limits and accountability. 

Why does this matter? 

What’s most exciting is that MHS could do for hardware what MCP has started to do for software.  

Once devices become easier for AI agents to discover and understand, experimentation becomes much more accessible. 

Whether that's optimising tomato plants in a greenhouse, investigating why a room is cold, accelerating scientific research, or coordinating climate monitoring at a global scale, the interesting part isn't just AI can control hardware, but that hardware has become easier to describe, share, connect, and collaborate with. 

And I suspect we're only just beginning to figure out what that might enable.