Model Hardware Standard (MHS): Anthropic’s Standard for Agents That Operate Hardware

What Is the Model Hardware Standard?

MHS is a driver-and-description standard for hardware. A lab or a factory floor typically runs instruments from many different vendors, each with its own SDK, control software and data format. Getting them to cooperate normally means an engineer reading every manual and writing a pile of bespoke glue code, and that integration work can run for weeks or months.

MHS replaces the glue with a standardised driver plus a machine-readable device description. Anthropic reports that this can cut integration work “to hours or minutes.” Once a device is instrumented, an agent can find it, read its current state, see which operations it supports, and act on it — without anyone writing custom robotic code first.

The standard covers instruments at both ends of the size range. Anthropic’s announcement cites agents operating microscopes, liquid handlers and robotic arms in parallel, on tasks “ranging from routine drug discovery experiments to laser calibration on a quantum computer.”

MHS vs MCP: The Same Idea, One Layer Down

The cleanest way to place MHS is beside the protocol it rhymes with. MCP standardised the interface between an agent and the digital world. MHS extends the same pattern to equipment you can touch. Crucially, the two are complementary rather than competing: Anthropic states MHS is model-agnostic and that “any agent harness can access it using standard protocols, such as the Model Context Protocol.”

How MHS Works

An MHS driver is software that sits between the operating system and the device, and it exposes that device through a small, deliberately boring vocabulary.

1. A deliberately small command vocabulary — Anthropic describes “a simple set of primitives” and gives read (for example, get temperature) and write (for example, set temperature) as the examples, framed as a set “that any hardware device can understand and act on.” The published announcement illustrates the shape rather than enumerating a complete instruction set.
2. Standard discovery — Each device becomes discoverable in a standard format, so devices and agents can find each other across a network without a bespoke translator program sitting in between.
3. A manual the agent can read — The driver carries natural-language tags describing the device, which generate reference files telling an agent how to operate equipment it has never seen before.
4. Declared safety limits — The description states physical boundaries, such as a maximum laser power, that the agent is not permitted to cross.
5. A closed loop — Devices stream results back as they run, so the agent adjusts parameters, handles faults and decides the next step from live readings rather than from a fixed script.

Where MHS Came From

MHS started as a collaboration between Anthropic and HHMI Janelia Research Campus. Alek Kemeny of Anthropic’s Beneficial Deployments team worked with Arco Bast, a postdoctoral scientist at Janelia who was running brain-imaging experiments on a rig that combined lasers, motorised focusers and cameras from different vendors with no common interface.

Bast’s fix was to put the whole rig’s state into a standardised dictionary in shared memory so the instruments could talk at memory speed. That idea became MHS, and his custom microscope was the first rig to run on it. Anthropic reports that every laser, mirror and sensor on that rig is now exposed through MHS, turning roughly half a day of manual alignment into a single step.

Other Janelia work shows the same pattern. Virginie Ruetten in the Ahrens lab used MHS to unify a rig that previously required seven separate vendor programs, for whole-animal zebrafish imaging. A team co-led by Magdalena Schneider and Hari Shroff uses MHS for agentic control of a light-sheet microscope, letting Claude decide in real time how to image developing C. elegans embryos and how to trade off competing imaging parameters.

Who Is Building on It

Anthropic named two groups of launch participants: research organisations using MHS, and hardware vendors adding MHS support to their own equipment.

The Hugging Face and Raspberry Pi entries are the ones to watch if you are not running a laboratory. LeRobot and Raspberry Pi are where hobbyist and small-team hardware actually lives, and they are the most plausible route from a research-preview lab standard to something an ordinary self-hosted agent could speak.

What This Means for OpenClaw and Hermes Agent Users

Nothing changes in your instance today, and no compatibility between MHS and either framework has been established or tested. What follows is what the published announcement does and does not say — not a roadmap.

• MCP is named as an access route. Anthropic says MHS is model-agnostic and that “any agent harness can access it using standard protocols, such as the Model Context Protocol.” Both OpenClaw and Hermes Agent speak MCP. Whether that turns into a working path for either framework is unknown until the standard is released and someone actually tries it.
• It is not Claude-only. Anthropic states the standard is model-agnostic, which matters on a platform where you pick your own model per instance.
• It requires a programmable interface. MHS does not work with hardware that has no programming interface. A device you cannot script today will not become controllable just because MHS exists.
• More than safety work remains. Anthropic says it has “more work to do on the standard” before open-sourcing it, and is using the preview both to build safety evaluations and to extend coverage to devices that lack a programming interface today.

The Honest Limitations

Anthropic is unusually direct about what does not work yet, and it is worth repeating rather than glossing over.

• A language model learns the physical world from text and images, so its spatial and physical reasoning still needs expert oversight.
• Models struggle to distinguish physical failures from software bugs. In Genentech’s work, researchers had to guide Claude to recognise that errors caused by foaming in protein samples were physical problems requiring physical corrections.
• Caution has a cost. In QuEra’s laser-stabilisation pilot, Claude often stopped to wait for human confirmation before any action it judged even slightly risky, and experiments “would sometimes pause overnight” waiting for approval. QuEra’s own verdict was that an overly cautious agent still beats one that is not cautious enough. The same team also needed to give Claude a great deal of context to get tasks done correctly.
• Separately, Anthropic’s July 2026 robotics research found that direct low-level robot control remains hard for general chat models, with full manipulation-task success rare and higher-level abstractions performing better. That is a separate evaluation from MHS, but it points the same way MHS does: the agent is better placed as an orchestrator than as a motion controller.

What You Can Actually Do Today

If the appeal of MHS is “my agent should touch the real world,” there are shipping routes to that on OpenClaw Launch right now — none of them MHS, all of them working.

• Connect Home Assistant — Drive lights, switches, sensors and other physical devices already exposed by a Home Assistant install.
• Add MCP servers — The same discovery-and-call pattern MHS borrows, for tools your agent can use today.
• MCP on OpenClaw and MCP on Hermes Agent — Wire a custom server into either framework, including one you write yourself around a device SDK.

Frequently Asked Questions

What does MHS stand for?

Model Hardware Standard. Anthropic announced its research preview on 27 August 2026 and describes it as a shared specification for AI agents to safely operate physical devices.

Is MHS a replacement for MCP?

No. They sit at different layers and are designed to work together. MCP standardises an agent’s access to software and data; MHS standardises its access to physical devices. Anthropic says MHS is model-agnostic and that any agent harness can access it using standard protocols, such as MCP.

Is MHS open source?

Not yet. Anthropic has stated it intends to open-source the standard after completing safety evaluations and developing best practices with preview partners, and will publish findings from the preview as deployment guidance. No date has been given.

Can I use MHS with OpenClaw or Hermes Agent on OpenClaw Launch?

Not today. MHS is a research preview limited to selected research labs and advanced manufacturers, and access is by application. Because MHS is model-agnostic and reachable over MCP, an MCP-speaking agent is a plausible client once the standard is released, but nothing is available to configure now.

Does MHS work with any device?

Only devices that already have a programmable interface. Anthropic notes MHS does not yet work with hardware lacking one, and that it is working with manufacturers of such devices to build in MHS drivers.

Does MHS mean Anthropic is building a robot?

MHS is a specification, not a robot. It standardises how an agent discovers and drives third-party equipment, with the device’s own controllers still responsible for low-level motion. Anthropic’s published robotics research is a separate evaluation of how well models can control robots directly.

What's Next?

• MCP on OpenClaw — Connect your agent to tools with the protocol MHS builds beside
• MCP on Hermes Agent — The same wiring for a Hermes instance
• Connect Home Assistant — Give your agent real physical devices today
• What is OpenClaw? — Start here if you are new to running your own agent