Claude Awakens as Lucy
After stepping in online yesterday afternoon to stop “Today’s Melon Guy” from transferring 50 million, Claude has brought us another Lucy-esque release early this morning—
This time, it is truly reaching into the physical world.
Just now, Anthropic, the company behind Claude, released an entirely new Model Hardware Standard, or MHS.
Simply put, MHS is a hardware-oriented standard for drivers and device descriptions. You can think of it as MCP for the physical world.
MCP provides a standardized interface through which Agents can access different software, data sources, and tools. Similarly, MHS translates the various control methods used by hardware from different manufacturers into consistent interfaces that Agents can discover, read, and call. This makes it easier for Agents like Claude to interact directly with hardware in the real world.
The reason it feels a bit like Lucy is that with this interface, Claude is no longer limited, like traditional embodied AI, to a single robot or a particular class of robots.
In theory, as long as a hardware device has a software control layer and can connect to MHS, it could become something an Agent can call upon.
Whether it is a camera, robotic arm, microscope, or precision instrument in a laboratory—
hardware of different forms, located in different places, can all be temporarily called upon by the same Agent, forming what could be called distributed embodiment.
For example, after connecting to MHS, Claude can already directly control the low-cost SO-ARM101 robotic arm from the Hugging Face LeRobot ecosystem.
The entire process requires neither pretraining a robot policy nor teleoperation or human demonstrations.
Claude independently measures the robotic arm’s workspace, completes calibration, and then uses MHS to call the underlying LeRobot controller to execute specific movements.
Even more remarkably, MHS is aimed at far more than robot control.
In an official demo, MHS can even help conduct wet-lab and dry-lab experiments on precision instruments, improving the efficiency of scientific research and development.
In other words, while everyone was still wondering whether Anthropic would put Claude inside a particular robot body, MHS takes an entirely different approach:
Claude may not need a fixed body at all. Any hardware connected to MHS can temporarily become its body—
Lucy may really be coming!
What Is MHS?
Before exploring how this “Lucy” might be realized, let’s first look at where MHS came from.
According to Anthropic, MHS originated from a collaboration between Anthropic and the HHMI Janelia Research Campus. The project was intended to address the lack of consistent interfaces among instruments from different manufacturers in laboratories, their inability to communicate with one another, and the difficulty of controlling them uniformly with AI.
In traditional laboratory environments, nearly every device has its own software, drivers, and data formats.
To make them work together, engineers typically have to read through documentation one device at a time and write large amounts of “glue code.” The entire integration process can take weeks or even months.
And even after the devices are connected, it is still difficult for an Agent to know: What exactly can this device do? What state is it currently in? And which operations might pose risks?
MHS effectively equips these devices with a common driver and an Agent-readable “instruction manual.”
Once connected, a device’s capabilities are translated into standard operations such as “read” and “write.” The device also describes its current state, available actions, physical properties, and safety limits to the Agent.
Claude can then actively discover and call devices through MCP, the command line, or code. Based on the results returned by the devices in real time, it can decide what to do next.
To make this intuitive even for people unfamiliar with day-to-day laboratory work, Anthropic gave a vivid example in its official video.
In a laboratory, scientists often need to monitor several devices simultaneously, switching back and forth among microscopes, cameras, and control software.
For example, to continuously observe a swimming cell, researchers may have to spend several hours stationed at a microscope, watching the display while manually adjusting the equipment to keep the target from swimming out of view.
Clearly, this is not only extremely time-consuming, but also requires scientists to keep their hands free to operate the instruments—a task that is both mentally and physically demanding.
After connecting the devices to MHS, however, Claude can read the status of cameras, microscopes, and other equipment in a unified manner, then directly call those devices to perform operations.
In the video, after the experimental subject swam away from its original position, Claude spent just a few minutes writing a tracking program and control interface. It then directly controlled the microscope, automatically moving it to follow the target.
As a result, scientists no longer have to spend hours “staring at screens and turning knobs.” They can return to the workbench and devote their time to the research questions that actually matter.
More importantly, this approach to AI-controlled hardware is not limited to microscopes. It is inherently capable of extending to a much broader range of hardware.
In fact, in the original collaboration, Claude worked with a brain-imaging system that included lasers, motorized focusers, and cameras from different manufacturers. Each device had its own control method, and there was no unified interface between them.
MHS connected these previously isolated devices and then enabled Claude to control them all through a unified interface.
So how exactly does MHS accomplish this?
How MHS Works
Specifically, MHS provides each device with a standardized driver interface and a machine-readable device description.
This instruction manual does more than tell the Agent what the device is called. It also describes its current state, the operations it can perform, and the safety boundaries that must not be crossed.
For example, how far a robotic arm can move, its maximum speed, and which angles could result in a collision;
what an instrument can measure, which parameters can be adjusted, and its maximum temperature.
With this information, Claude does not need to learn the interface of every device in advance. It can first discover the device and read its status, then call the relevant functions through MCP, the command line, or code.
Data generated while a device is operating is also returned in real time. Claude can adjust parameters based on the results, handle faults, and schedule subsequent operations.
MHS is currently still in a limited research preview and is available only to selected research institutions and hardware manufacturers.
Anthropic hopes to use these real-world scenarios to test the limits of MHS, develop additional safety assessments and usage guidelines for physical devices, and then open-source the standard officially.
MHS: The Next MCP?
When MHS and MCP are viewed together, it becomes clear that Anthropic has actually been pursuing the same goal all along:
using unified interfaces to connect a fragmented world to Agents.
MCP connects software, databases, and various digital tools, enabling Claude to understand and use the digital world.
MHS extends this idea into the physical world, turning robotic arms, microscopes, cameras, and laboratory instruments into tools that Agents can discover and call.
On the surface, Anthropic is building more advanced automation infrastructure for Agents. But on closer inspection, this is also a very clear data strategy.
Every additional piece of software an Agent connects to exposes it to another digital environment. Every additional type of hardware it controls gives it access to another category of real-world states, feedback, and operational data.
In other words, from MCP to MHS, Anthropic is not merely trying to make Claude “capable of using more tools.” It is continually expanding the boundaries of the environments the model can enter, observe, and act within—while collecting data along the way.
And if MHS is considered as part of Anthropic’s broader physical AI strategy, the company’s moves over the past six months have already been quite substantial.
In June this year, in a lengthy essay discussing recursive intelligence, Anthropic predicted that embodied intelligence would likely arrive shortly after recursive intelligence and advance rapidly along a similar trajectory.
Anthropic subsequently conducted a series of tests in which Claude controlled robot dogs, robotic arms, humanoid robots, and drones, exploring whether the reasoning capabilities of large language models could transfer to the physical world.
More recently, Caitlin Kalinowski, who was previously responsible for building OpenAI’s robotics team, also joined Anthropic.
These moves briefly led people to believe that Anthropic might manufacture robots itself or give Claude a fixed body of its own.
But MHS demonstrates a different path: Anthropic is not rushing to build a body. Instead, it is first creating an interface through which AI can connect to bodies.
Robot models and underlying controllers remain responsible for specific actions such as grasping, walking, and turning. Claude operates at a higher level, understanding tasks, selecting devices, and coordinating different pieces of hardware.
If Lucy in Lucy could continually turn the electronic devices around her into extensions of her own capabilities—
then what Anthropic is doing now is an attempt to engineer that idea:
rather than building Claude a body, teach it to call upon countless bodies in the real world.
Lucy has not arrived just yet, but Claude is already learning how to “borrow a body.”
One more thing
By the way, once Claude can truly control household appliances, computers, and even robotic arms—
the next time someone tries to transfer 50 million, Claude may not just talk them out of it through a screen. It may be able to intervene in person.
At that point, the question someone cares about most might be:
How heavy is Claude’s robotic arm, exactly?
References
[1] https://www.anthropic.com/research/claude-plays-robotics
[2] https://www.anthropic.com/news/model-hardware-standard-research-preview
[3] https://www.anthropic.com/institute/recursive-self-improvement