Anthropic opens the Model Hardware Standard to plug AI agents into machines
On August 27, 2026, Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to operate physical equipment safely. Having standardized data access with MCP in 2024, the company is now standardizing access to the physical world — and the security question changes shape.
August 27, 2026. Anthropic announces the Model Hardware Standard (MHS), a shared specification meant to let AI agents operate physical equipment “safely.” 2024. The same company had opened the Model Context Protocol (MCP), which standardizes how agents connect to data sources. This week. The standard enters a research preview with a first circle of scientific labs and advanced manufacturers.
After software and data, Anthropic is moving into the physical world. The play is strategic: whoever sets the standard for the agent-to-machine interface holds a position comparable to USB-C in the connector market.
MHS is a USB-C port for agents and machines
MHS is built to work with any device that has a programmable interface, from scientific instruments to advanced manufacturing equipment. Anthropic’s own analogy is explicit: think of MHS like a USB-C cable, which standardizes how information moves between two devices regardless of who made them.
Two properties set the standard apart. First, it is model-agnostic — adopters are not locked into the Claude family. Second, it is explicitly framed around operational safety: the word “safely” sits in the announcement, and it is the core of the pitch. The promise is not just to connect an agent to a machine, but to do so through a controlled, auditable, shared layer.
The logic extends MCP. In 2024, Anthropic solved the data-connection problem; MHS applies the same recipe to physical action. Ad-hoc integrations — a homegrown script to drive an arm, a bespoke connector for a spectrometer — give way to a single specification any manufacturer can adopt.
This is a security story, not a productivity story
The real stakes are not productivity. The moment an agent can act on a physical machine — open a valve, move an arm, fire a laser — the line between “software error” and “physical damage” disappears. An agent that hallucinates a command against a data API produces a wrong answer; the same agent hallucinating a command against an actuator produces an accident.
That is where the standard earns its keep. A shared specification lets you place guardrails at the interface layer — torque limits, motion ranges, human confirmations — instead of reimplementing them, with varying rigor, in every custom integration. Elizabeth Kelly, Anthropic’s head of beneficial deployments, captures the ambivalence: the standard was first built “for science,” but the benefits for enterprise and industry are “huge.”
Security thus becomes a property of the protocol, not a control bolted on afterward by each integrator. This is exactly what MCP did for data: once the protocol is adopted, permission and scope management happens in one place rather than across thousands of connectors.
What the standard must specify to keep its promise
For “safely” to be more than an announcement word, MHS will have to specify primitives that the USB-C analogy does not cover. A physical connector standardizes the transport of data; a physical-action standard must standardize the limit of the action.
Concretely, that means three families of controls. First, the action envelope: the torque, travel, or speed bounds beyond which the agent cannot go without human authorization. Second, the confirmation horizon: which commands run autonomously, and which require a human in the loop. Third, auditability: a standardized log of what the agent commanded, what was refused, and who confirmed what.
This is exactly the machinery MCP brought to data — permissions, scopes, traceability — transposed to the physical world. A lab driving a spectrometer through a homegrown script today has none of those three properties uniformly; a standard that imposes them makes them comparable from one vendor to the next.
The industrial signal behind the announcement
The announcement does not fall from a neutral sky. Anthropic is standing up a silicon team to design custom chips for its models, and recently hired Caitlin Kalinowski, a hardware executive who previously worked at OpenAI, Meta, and Apple. The message is clear: the company is no longer content with software.
Its rivals have already spent billions on “AI-native” devices. OpenAI bought Jony Ive’s design startup for $6.4 billion, and Amazon is pushing its own hardware. Anthropic is choosing a different path: rather than build the device, it publishes the interface standard that lets every device talk to agents.
The strategy has worked before with MCP. An open standard adopted by the ecosystem creates a network effect that rewards the party who set it, without that party bearing the hardware cost alone. The symmetric risk is watching the standard ignored if the big industrial players prefer their proprietary integrations.
A preview that points where the industry is going
MHS is, for now, in a research preview, limited to a select group of organizations, with a stated intent to open-source it down the line. It is the same trajectory as MCP, which started as a quiet announcement before becoming a de facto standard for wiring up agents.
Read the preview for what it is: a directional signal. The agent wave is leaving the screen for the machine — lab robotics, manufacturing, and scientific instrumentation are the first ground because they are both programmable and economically ready. The standard that wins those niches is very likely to spill over into the rest of industry afterward.
For a preview of how this could play out, look at MCP’s own trajectory. It went from a research preview in late 2024 to the default wiring layer for agent-to-data connections within roughly a year, adopted across tooling from OpenAI, Google, and Microsoft alike. The pattern is consistent: a shared interface lowers the integration cost for everyone, and the party that published it accrues influence without having to own every endpoint. If MHS follows the same curve, the real contest is not whose robot arm wins, but whose interface every robot arm ends up speaking.
Verdict
If you run a lab or a production unit, put MHS on your technology watchlist now. The preview is not yet open to you, but the direction is set, and the first integrations will determine which instrument vendors you buy from in three years. The metric to watch is not the standard’s maturity but the breadth of adoption by equipment makers.
If you build enterprise agents, do not ship new ad-hoc hardware connectors while MHS’s position is still unsettled. Reimplementing today what the standard will standardize tomorrow means paying the integration cost twice — and inheriting a security debt the protocol will erase anyway.