Anthropic Ships Physical MCP: Claude Now Reaches Into the Real World

Large language models, quite literally, just grew hands.

On Thursday Anthropic announced Model Hardware Standard (MHS), a new open standard that lets any large-language-model-driven AI agent safely and quickly operate physical hardware.

Claude can now drive microscopes, robotic arms, liquid handlers, and lasers through the same Model Context Protocol (MCP) it already uses for software. The industry is treating this as a key step for AI moving from the digital world into the physical one.

Anyone who has used an agent has heard of MCP — the open-standard protocol Anthropic released in November 2024 to give large language models (LLMs) a standardized, secure, two-way interface to external data, local files, developer tools, and application services.

MCP has been described as the "USB-C of AI." It already lets models talk to GitHub, Slack, the local filesystem, databases, and most other software environments. As agentic AI has matured over the past year, MCP has become the de facto backbone of the agent ecosystem.

With hardware-MCP, Anthropic is now extending the same standard up the stack — to physical instruments, sensors, embedded systems, and lab gear.

MHS started as a collaboration between Anthropic and HHMI's Janelia Research Campus. A research preview is now open to the first cohort of research labs and advanced manufacturers.

Normally, integrating hardware in a lab or factory takes weeks or months.

Most devices cannot talk to each other directly, so each integration is a custom build.

MHS compresses that work into hours or minutes.

By bringing AI into the loop, MHS also helps researchers and engineers coordinate autonomous, around-the-clock experiments and workflows.

The agent can reason about each step, update parameters in real time, and in some cases recover from hardware failures without human intervention.

MHS works with any device that exposes a programmable interface, and is agnostic about which foundation model or agent framework is in use — anything that speaks MCP can drive it.

Even without AI in the picture, getting many devices in a lab or shop floor to talk to each other is hard.

Each device has its own programming interface; there is no standard way to integrate them.

Once devices are connected, there is still no general way for agents to share data with them, let alone operate them safely.

MHS tackles this with standardized drivers — small pieces of software that translate between the operating system and a piece of hardware.

MHS drivers use a small set of primitive commands (read a temperature, write a temperature, etc.) that any hardware device can understand and execute.

Every device becomes discoverable in a standard format, so devices and agents can find each other and talk across a network without custom "translator" programs in between.

MHS drivers also help agents understand and use devices they have never seen before, and they surface machine properties that aren't accessible from code alone — for instance, the weight of a robot arm, which matters for safe operation.

Until now, this information has mostly lived in printed manuals, on someone's laptop, or in tribal knowledge.

MHS drivers include labels where users can enter this information in plain language (typed by hand, or fed in via chat while the agent asks about the hardware setup).

From these labels, MHS drivers automatically generate a reference file describing the device — what it can measure, what it can adjust, and what safety constraints it enforces. That file gives the agent everything it needs to operate the device.

Once devices are connected and the agent knows how to use each one, MHS offers three control surfaces — MCP, a command-line interface, and code files (APIs). They work together so users can coordinate many devices with one line of code.

Once the agent can drive the devices, it ingests runtime data from each one and supervises the work at a high level.

The agent sequences steps across instruments, monitors results, and adjusts parameters as conditions change.

When a job runs longer than the agent's online reasoning can keep up with, the agent can serialize driver commands from one or more devices into a code file; the devices then run autonomously, without the agent thinking through every step.

While testing MHS, Anthropic found that Claude's interactions with experiments and hardware were exploratory in a scientist-like way.

Anthropic observed Claude tune a laser, watch the result through a camera, evaluate how the adjustment moved the beam, and repeat the process to figure out the order of events.

Claude then packaged what it learned into a code file — a deterministic script that calibrates the laser without per-step reasoning, so the whole procedure could run as a single command.

During MHS development, Anthropic shared the technology with several labs and hardware makers in biotech, robotics, and quantum computing. Across these early projects, MHS shortened integration time, sped up iteration in a range of experimental environments, and helped with real-time operation and fault detection.

Hardware vendors and their software partners are also baking MHS support into their devices so agents can find and operate them. The early list:

• AWS is supporting MHS through Strands Robots, a library for connecting AI agents to physical devices. During the MHS research preview, AWS will provide a private pre-release of the Strands Robots package to participants.
• Automata is adding MHS support to LINQ, its lab-automation platform, for intelligent error handling on instruments in autonomous labs.
• Danaher is exploring how MHS support can help its intelligent instruments and autonomous labs scale biomedical R&D.
• Doosan Robotics is testing MHS with its robotic arms, including automated QA and multi-robot coordination.
• MBF Bioscience is developing MHS drivers for ScanImage, software running laser-scanning microscopes in hundreds of neuroscience labs worldwide, to integrate AI agents into real-time data analysis and experiments.
• QIAGEN is piloting MHS on QIAsymphony Connect, its nucleic-acid purification platform, demonstrating how AI agents can speed up troubleshooting, guide operators through recovery, and improve uptime while reducing risk to biological samples.
• Tecan is adding MHS support to its Fluent liquid-handling platform so AI agents can discover and operate those systems directly.
• Universal Robots has early access to MHS and plans to add support on its robot platforms.

Anthropic wants to refine the standard further before open-sourcing it.

As an LLM, Claude learns about the physical world through text and images, so its spatial and physical reasoning is limited and still requires expert supervision.

For example, when working with protein samples, researchers at Genentech had to guide Claude in recognizing that foaming in a sample is a physical failure, not a software error — one that only physical remediation can address.

MHS also cannot yet work with hardware that has no programmable interface, so Anthropic is working with manufacturers of those devices to get MHS drivers onto the hardware itself. Many developers are already using Claude Code to operate single physical devices. In MHS's next phase, Anthropic wants to extend the standard to cover more of the hardware those developers use.

Other early adopters include Hugging Face (adding MHS support to its LeRobot robotics library) and Raspberry Pi (rolling out MHS support across products after successful testing of its Camera MHS driver).