Anthropic推出硬件标准MHS：让AI Agent统一控制机械臂和实验室仪器

Anthropic is building the Model Hardware Standard (MHS), a unified interface that lets AI agents read data from and control physical devices like microscopes and robotic arms in labs and factories.

The company uses standardized drivers that make every device discoverable in a common format. According to Anthropic, this cuts integration time for connecting multiple machines from weeks or months down to hours or minutes.

Early tests with partners show that AI can independently optimize workflows and generate scripts that run without a language model. But the system still struggles with understanding physical cause and effect, so human oversight remains a must.

Anthropic's Model Hardware Standard (MHS) gives AI agents a unified interface to physical devices like microscopes and robotic arms. Early tests in labs and on quantum computers cut integration time dramatically, though Claude still struggled with physical cause and effect.

Devices in research labs and factories often come from different manufacturers, use different APIs, data formats, and control software, and take serious effort to connect. Anthropic wants to tackle this with the Model Hardware Standard (MHS), a spec that lets AI agents read from and control physical hardware.

Anthropic says MHS cuts integration time for lab equipment from the typical weeks or months down to hours or minutes. The company developed the spec with the HHMI Janelia Research Campus and is releasing it first as a research preview for select labs and manufacturers, with an open-source release planned later.

One driver per device instead of custom integrations

The idea builds on a project Anthropic already shipped for software. The Model Context Protocol (MCP) gives AI models a standard way to access external data sources and tools. MHS extends that approach to machines.

Each device gets an MHS driver that unifies basic functions like reading and modifying data. The driver makes the device discoverable in a common format and can include details software alone can't capture, such as a robotic arm's weight or its safety limits. Users can add this information in natural language, and MHS turns it into a reference file for the agent.

An agent can then pick up the capabilities and limits of a device it has never seen before and operate it through a single interface. Someone still has to write a driver for each piece of hardware, but once it exists, it can be reused rather than rebuilt for every combination of device and controller.

According to Anthropic, MHS is model-agnostic and works with any device that has a programmable interface. Agents can coordinate multiple devices and save workflows as conventional scripts that then run without a language model.

How Claude handled real lab equipment

Anthropic tested MHS with several partners. At biotech company Genentech, Claude coordinated a liquid handler, a robotic arm, and a plate reader to automate a protein assay, a test that measures protein concentration. The agent optimized pipetting parameters for different liquids on its own, but ran into trouble when bubbles formed in a viscous solution and caused errors. Claude kept restarting the process in the same vessel with tweaked parameters, which made things worse. It took a human explaining the physical cause before the agent found a workable fix.

At Carnegie Mellon University, researchers connected a liquid handler, a plate reader, a robotic arm, and several monitoring cameras across three computers with fundamentally incompatible interfaces. The team says the drivers and orchestration layer took about eight hours to build, compared to the several weeks a vendor setup would typically require. Claude also noticed that a first experimental run hadn't produced a good enough measurement curve, lowered the maximum concentration of the substance being tested, and got a much better result on the second try.

At quantum computing company QuEra, Claude used MHS across hundreds of automated runs to develop a control program that brings a laser back to a stable operating state after disruptions. The resulting script succeeded in 695 out of 700 attempts in a blind test, a 99.3 percent success rate, running entirely on its own with the language model out of the loop.

All three examples come from partner projects presented by Anthropic and haven't been independently verified.

Expert oversight is still part of the setup

Because Claude learns about the physical world through text and images, its spatial and physical reasoning has limits that still require expert oversight, Anthropic says. In the Genentech tests, researchers had to guide Claude to recognize that errors caused by foaming samples were physical failures rather than software bugs. During the research preview, the company plans to build additional safety evaluations with its partners and is developing a physical safety roadmap to guard against misuse.

If the approach holds up, it could make it much easier for AI agents to move beyond purely digital tasks into labs, factories, and other physical systems. Several manufacturers and organizations are already building MHS support or testing the spec, including AWS, Doosan Robotics, QIAGEN, Tecan, Universal Robots, Hugging Face, and Raspberry Pi.