On August 27, 2026, Anthropic announced the Model Hardware Standard (MHS) — an open specification that does for physical machines what the Model Context Protocol (MCP) did for software tools. AI agents can now read sensors, write to actuators, and operate physical equipment like robotic arms, microscopes, liquid handlers, and lasers directly through a standard interface.

The announcement is in research preview, with an initial cohort of scientific labs and advanced manufacturers. Anthropic plans to open source the full specification.

What MHS Actually Does

If you have been following AI agent development, you know MCP. It is the standard that lets AI agents connect to databases, APIs, and software tools through a shared protocol instead of custom point-to-point integrations. MHS is the same concept extended into the physical world.

Before MHS, connecting an AI agent to a piece of lab equipment or factory machinery required custom driver code for every device, often taking weeks or months to build and test. MHS cuts that to hours by providing a standard reference file for each device that tells the agent three things: what the machine can measure, what it can change, and which safety limits it must respect.

The protocol adds three capabilities that MCP does not have because it was built for software:

Physical control primitives. MHS defines read and write operations for sensors and actuators — the fundamental building blocks of any hardware interface. This is not a high-level API; it is a protocol that speaks the language physical devices understand.

Device-level safety limits. The specification includes constraints that block operations before the machinery moves. Anthropic gave specific examples: preventing collisions during robotic arm transfers, blocking laser power settings that would damage samples, detecting missing or rotated plates, and triggering emergency stops when anomalies are detected. Safety is baked into the protocol, not bolted on as an afterthought.

Automatic network discovery. Agents can discover what hardware is connected to a network without manual configuration. This matters for labs and factories where equipment inventories change frequently.

Who This Is Built For Right Now

The research preview is targeted at scientific research labs and advanced manufacturers — environments where physical automation has real economic value and where the stakes of a mistake are high enough to justify careful rollout.

Scientists who spend hours manually operating liquid handlers, microscopes, and measurement equipment are the clearest immediate beneficiaries. An AI agent that can design an experiment, operate the equipment, collect the data, and analyze the results without a researcher physically at the bench is a significant capability jump.

For manufacturers, the value proposition is predictable: equipment that operates reliably under AI direction, with safety constraints that prevent costly errors, reduces the labor cost of running complex machinery and catches errors that tired human operators miss.

Model-Agnostic by Design

MHS works with any language model — not just Claude. The specification is designed to be adopted by the broader AI ecosystem, which is why open sourcing it is part of the plan. Anthropic is positioning this as an industry standard, not a proprietary advantage.

This mirrors the trajectory of MCP, which started as an Anthropic project and has since been adopted across the AI industry. If MHS follows the same path, it becomes the standard interface between AI agents and physical hardware regardless of which model an organization runs.

What This Means for Business

The transition from software-only AI agents to physical AI agents is a meaningful shift. Most of the conversation about AI agents in enterprise settings has focused on automating digital workflows: reading emails, updating CRM records, generating reports, filing documents. MHS points toward the next frontier: agents that can also operate the physical processes those digital workflows are connected to.

For businesses in manufacturing, life sciences, agriculture, energy, and logistics, this is where AI automation gets genuinely disruptive. The ability to run physical processes under AI direction, with real safety constraints, changes the economics of both labor and quality control.

A few practical questions worth tracking as this develops:

Integration complexity. MHS is still in research preview. The standard will evolve based on what early adopters learn. Businesses considering physical AI automation should watch how the specification develops before building around it.

Liability and governance. When an AI agent operates a piece of manufacturing equipment and something goes wrong, who is responsible? The safety limits in MHS are a start, but the governance frameworks for physical AI automation are still being written, by courts and regulators as much as by engineers.

Device coverage. The initial preview focuses on scientific lab equipment and some manufacturing applications. Broader device support will expand the addressable market, but that takes time.

Workforce implications. AI agents operating physical equipment will change the nature of operator roles in the industries that adopt it earliest. This is not a distant concern for businesses in manufacturing and logistics.

The most important near-term takeaway for business leaders: the scope of what AI agents can automate just got significantly larger. The AI agent workforce is no longer limited to the digital side of your operations.

Enterprise DNA is watching MHS closely. Businesses considering AI agent deployment as part of their operational strategy should factor physical AI capabilities into their planning now. For help thinking through what AI automation means for your specific operations, Omni Advisory works with business leaders on exactly that question.