Claude Takes Over the Physical World: Anthropic Launches Model Hardware Standard

2026-09-02 · AI Tech Insights · Curated from 量子位

Summary: Anthropic released the Model Hardware Standard (MHS), dubbed 'MCP for the physical world.' It translates the control interfaces of robotic arms, microscopes, and cameras into agent-discoverable, agent-callable APIs—usable with zero pre-programming. This enables a new 'distributed embodiment' paradigm: Claude need not own a fixed body; any MHS-compatible hardware can temporarily become its body.

Image source: 量子位

Anthropic yesterday released the Model Hardware Standard (MHS)—a set of hardware-oriented drivers and device description standards. Its design goal is straightforward: translate the proprietary control interfaces of different vendors' hardware into a consistent interface that agents can discover, read, and invoke—so that agents like Claude can directly interact with real-world hardware.

MHS's core idea inherits from MCP (Model Context Protocol). If MCP unifies software, data sources, and tools into a standard interface callable by agents, MHS extends the same idea to the physical world—cameras, robotic arms, microscopes, lab instruments—as long as they conform to MHS, any of them can be temporarily invoked by the same agent, forming what Anthropic calls "distributed embodiment."

The key demonstration is with Hugging Face LeRobot's low-cost SO-ARM101 robotic arm. After MHS integration, Claude controls the arm end-to-end—no pre-trained policy, no teleoperation, no human demo. Claude measures the workspace, completes calibration, then uses MHS to call LeRobot's underlying controllers. In other words, agent control of physical machines has, for the first time, a standardized interface like MCP.

The more imaginative application is in the lab. Anthropic demoed a microscope auto-tracking a swimming cell: after the specimen drifted out of view, Claude wrote a tracking program and UI in minutes, controlling the microscope to follow the target automatically. No scientist needs to stare at a screen or turn a knob—they can focus on the real research question.

MHS originated from a collaboration between Anthropic and HHMI Janelia Research Campus. In traditional labs, each device has its own software, drivers, and data formats. To make them collaborate, engineers typically have to read each manual and write extensive glue code—integration cycles measured in weeks or even months. MHS provides each device with a standardized driver interface and a machine-readable device description: this "manual" tells the agent what the device is called, its current state, what actions it can perform, and which safety boundaries cannot be crossed.

Notably, MHS targets far more than robots. The official demo also includes dry/wet experiments with precision instruments. While people were still guessing whether Anthropic would put Claude inside a robot's body, MHS takes a different approach: Claude doesn't need a fixed body—any MHS-compatible hardware can temporarily become its body.

If Lucy in the movie Lucy could continually turn surrounding electronics into extensions of herself, Anthropic is engineering exactly that: instead of building Claude a body, teach it to call countless bodies in the real world. The movie Lucy hasn't arrived, but Claude has already started "borrowing bodies."

MHS is currently in limited research preview, available only to select research institutions and hardware vendors. Anthropic plans to test the capability boundary in real scenarios, supplement safety evaluations and usage norms, then open-source the standard formally.

Key Takeaways

• MHS is billed as 'MCP for the physical world' — a standardized hardware interface
• Agents can control the SO-ARM101 arm with no pre-trained policy or teleoperation
• Enables 'distributed embodiment': Claude needs no fixed body of its own
• Demo scenario: microscope auto-tracks a swimming cell; tracking code written in minutes
• MHS is in limited research preview, with open-sourcing planned

📎 This article was automatically compiled by AI from 量子位 (2026-08-28).
All rights belong to the original authors. Used for informational purposes only.