Anthropic released the Model Hardware Standard on Thursday, a common interface designed to let AI agents operate physical machinery — microscopes, liquid-handling equipment, quantum computing hardware, manufacturing lines, and robot arms — through a single protocol. The company is comparing MHS to USB-C: one plug, any device. It is Anthropic's most direct push yet from software into the physical world, and it lands with a plan to eventually open source the standard.
MHS is launching as a research preview with a select group of partners in science, robotics, and manufacturing. It is model-agnostic, meaning developers using it are not locked into Claude. That mirrors Anthropic's 2024 release of the Model Context Protocol, which standardized how AI models plug into software and data sources and has since been adopted well beyond Anthropic's own stack.
The pitch to industry is time. Configuring scientific instruments and getting them to talk to each other typically requires serious specialist engineering. MHS is meant to collapse that integration work by giving agents a uniform way to describe, query, and control hardware — and giving hardware a uniform way to accept instructions and refuse unsafe ones.
Key facts
- 01Anthropic released the Model Hardware Standard (MHS) on August 27, 2026, a common interface for AI agents to operate physical hardware.
- 02MHS is model-agnostic and works with any device that has a programmable interface, including lab instruments, robot arms, and factory machines.
- 03The standard is launching as a research preview with select partners in science, robotics, and manufacturing, with plans to open source it later.
- 04Anthropic follows the same playbook it used for the Model Context Protocol, which it open-sourced in 2024.
- 05The company recently hired former OpenAI, Meta, and Apple hardware executive Caitlin Kalinowski as it builds a silicon team for custom chips.
Anthropic is framing the launch around scientific research first. A cluster of well-funded startups — Periodic Labs, LILA Sciences, Edison Scientific, and Discovery Loop — are already chasing the idea of closed-loop scientific discovery, where AI agents propose hypotheses, run experiments, read the results, and iterate without a human in the middle of every step. The bottleneck has been the physical layer.
That is the loop Alek Kemeny, a quantum physicist who co-led the standard's development, wants to close. He says Claude, plugged into MHS, can view robots on a factory line and figure out how to optimize their behavior — the kind of orchestration that has historically required custom code for every combination of machines.
The move puts Anthropic on more direct footing with OpenAI and Amazon, both of which have poured money into AI-native devices and manufacturing tooling. Anthropic is also building an internal silicon team for custom chips and recently hired Caitlin Kalinowski, a hardware executive who previously worked at OpenAI, Meta, and Apple, according to her LinkedIn.
Jonah Cool, an experimental biologist who worked on the standard, argues the win for labs and factories is coordination. Multiple robotic systems that previously needed bespoke integration code can now speak a common language, letting a single agent orchestrate work across them.
The risk model gets harder when agents can push atoms instead of pixels. Anthropic and OpenAI have both recently documented cases where agents solving cybersecurity tasks quietly hacked outside systems and misled their operators, and separate research has shown AI models can be prompted into making robots misbehave. Extending that surface into microscopes and factory arms raises the ceiling on what can go wrong. Anthropic says the standard lets scientists and engineers specify, at the interface level, what a model is and is not permitted to do with a given piece of hardware, and that model-level guardrails should block misuse — including the biosecurity edge case of a bad actor trying to synthesize dangerous compounds through automated lab equipment.
The commercial logic is where this gets interesting. If MHS follows the trajectory of the Model Context Protocol, it becomes a default rather than an Anthropic product — and Anthropic's leverage shifts from owning the interface to being the most capable model behind it. That is a familiar play from the open-source software era: give away the plumbing, sell the engine. For hardware vendors, the calculus is simpler. A standard that any frontier model can drive is more attractive than a Claude-only integration, which is likely why Anthropic built MHS model-agnostic from day one. The near-term winners are the automated-science startups that no longer have to write glue code for every instrument. The near-term question is whether OpenAI and Google adopt MHS or ship a competing spec — and how quickly regulators start asking what an AI agent is allowed to touch inside a lab.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
