Large language models have truly grown hands now.
This Thursday, Anthropic announced the launch of MHS (Model Hardware Standard), a new standard designed to help all large model-powered AI agents safely and efficiently control physical devices.
Now, large models like Claude can operate hardware such as microscopes, robotic arms, liquid handlers, and lasers in the physical world using MCP (Model Context Protocol). This advancement is regarded by the industry as a crucial step in AI’s transition from the digital world to the physical world.
Anyone using agents has likely heard of MCP, an open standard protocol introduced by Anthropic in November 2024, designed to provide a standardized, secure bidirectional communication interface between large language models (LLMs) and external data sources, local files, development tools, and various application services.
MCP is fundamentally positioned as the "USB-C interface for AI," enabling seamless integration between large models and software environments such as GitHub, Slack, local file systems, and databases. With the advancement of AI technologies in recent years, MCP has evolved into an indispensable de facto standard within the agent ecosystem.
Today, Anthropic is also leveraging hardware to implement MCP, elevating this standardized communication protocol to physical hardware, sensors, embedded systems, and test equipment.
The development of MHS began with a collaboration between Anthropic and the Howard Hughes Medical Institute's Janelia Research Campus, and the research preview is now available to首批 scientific laboratories and advanced manufacturers.
Typically, laboratories or manufacturing facilities require weeks or even months to set up and integrate hardware. Since most devices cannot communicate directly, experts must build customized integration solutions. MHS reduces this integration time to just hours or even minutes. Additionally, by integrating AI into these tools, MHS helps researchers and engineers more easily coordinate autonomous, 24/7 experiments and workflows. Agents can reason through each step of an experiment, update parameters in real time, and in some cases, recover from hardware failures without human intervention.
MHS can be used with any device that has a programmable interface, regardless of the type of base model, and any agent framework can access it using standard protocols such as MCP.
Even without considering the additional complexity of integrating AI into the system, enabling multiple devices in a laboratory or factory floor to communicate with each other is no easy task. Each device typically has its own programming interface, and there is currently no standardized method for integration. Furthermore, once devices are connected, there is no universal way for them to share data with agents, let alone for agents to securely operate these devices.
Now, MHS addresses these challenges by introducing a standardized driver: a software component that translates between a computer’s operating system and hardware devices. The MHS driver uses a set of simple, fundamental commands—such as read (e.g., retrieve temperature) and write (e.g., set temperature)—that any hardware device can understand and execute. This enables each device to be discovered in a standard format, allowing devices and agents to discover each other and communicate across the network without requiring custom “translation” programs between them.
MHS drivers also help agents understand and use previously unseen devices, providing machine-specific information that may not be obtainable through code alone—such as the weight of a robotic arm, which is critical for safe operation. Until now, this information has mostly been stored in paper manuals, user computers, or as tacit knowledge. But MHS drivers include tags that allow users to input this information directly using natural language—either by entering it themselves or through a conversation with the agent, where the AI queries their hardware setup.
Using the information from these tags, the MHS driver automatically generates a reference file containing general device characteristics, such as what it can measure, what can be adjusted, and which security restrictions will be enforced. This file provides the agent with all the information needed to operate the device.
Once the devices are connected and the agent program understands how to use each device, a method is still needed to control this hardware. For MHS, there are three such mechanisms: MCP, the command-line interface, and code files (API). They work together to enable users to coordinate multiple devices with a single line of code.
Once the agent can control the devices, it can receive operational data from each device and oversee and guide operations at a high level. The agent can sequence operational steps for different instruments, monitor results, and adjust parameters in response to real-time changes. When the agent needs to execute long-running tasks or operate devices at a speed exceeding its online reasoning capacity, it can chain driver commands from one or more devices into code files. This allows the devices to perform operations autonomously without requiring the agent to reason through each step.
During testing of MHS, Anthropic found that Claude interacted with experiments and hardware in a highly exploratory, scientist-like manner.
Anthropic observed that Claude adjusted the laser and used a camera to observe the results, evaluating how the adjustments moved the laser beam, then repeated this process to understand the sequence of events. Afterwards, Claude packaged the learned information into a code file, writing a deterministic script that could calibrate the laser without reasoning through each step, allowing the entire process to be executed as a single command.
During the development of MHS, Anthropic shared this technology with several laboratories and hardware manufacturers in fields such as biotechnology, robotics, and quantum computing. In these early projects, we observed that MHS reduced device integration time, accelerated iteration speeds across various experimental environments, and enabled real-time machine operation and fault detection.
Hardware vendors and their accompanying software companies have also integrated MHS support into their devices to enable agents to discover and operate them. For example:
Amazon Web Services (AWS) supports MHS through Strands Robots, a library for connecting AI agents to physical devices. During the MHS research preview, AWS will provide participants with a private pre-release version of the Strands Robots software package.
Automata is adding MHS support to its laboratory automation platform, LINQ, to enable intelligent error handling for instruments in autonomous labs.
Danaher is actively exploring how MHS-supported features can enable its smart instruments and autonomous labs to scale biomedical research and development.
Doosan Robotics is using its robotic arms to test MHS, including performing automated quality assurance and coordinating tasks among multiple robots.
MBF Bioscience is developing an MHS driver for ScanImage, the software used to operate laser scanning microscopes in hundreds of neuroscience laboratories worldwide, with the goal of integrating AI agents into real-time data analysis and experiments.
QIAGEN is testing MHS through a proof-of-concept on its nucleic acid purification platform, QIAsymphony Connect, to demonstrate how AI agents can help laboratories faster troubleshoot instrument issues, guide operators through recovery procedures, increase instrument uptime, and reduce risks to biological samples.
Tecan is adding MHS support to its Fluent liquid handling platform so AI agents can directly discover and operate these systems.
Universal Robots has obtained early access to MHS and plans to add support for MHS on its robotic platform.
Before open-sourcing, Anthropic still wants to further refine this standard.
As a large language model, Claude learns about the physical world through text and images, which means its spatial and physical reasoning capabilities are limited and still require expert oversight. For example, when processing protein samples, Genentech researchers had to guide Claude to recognize that bubbles in the sample indicated a physical malfunction rather than a software error—something that could only be resolved through appropriate physical corrections.
MHS is currently incompatible with hardware lacking a programming interface, so Anthropic is collaborating with manufacturers of such devices to integrate MHS drivers directly into the hardware. Many developers are already using Claude Code to operate individual physical devices. In the next phase of MHS, Anthropic aims to expand this standard to support a broader range of devices used by developers.
Early adopters include Hugging Face, which is adding MHS support to its robot library LeRobot, and Raspberry Pi, which is enabling MHS integration across multiple products after successfully testing with its Camera MHS driver.
Reference content:
https://www.anthropic.com/news/model-hardware-standard-research-preview
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by Machine Heart, focused on agents.