Anthropic Previews Model Hardware Standard for AI-Controlled Lab and Factory Equipment
Anthropic’s Model Hardware Standard gives AI agents a common, safety-aware interface for operating scientific and manufacturing equipment.
Anthropic has opened the first research-preview phase of the Model Hardware Standard, or MHS, a shared specification designed to let AI agents operate physical equipment through a common, safety-aware interface.
The preview covers laboratory and advanced-manufacturing systems including microscopes, liquid handlers, robotic arms and quantum-computing hardware. Access is currently limited to selected partners and applicants. Anthropic says it intends to release MHS as open source after using the preview to develop safety evaluations and deployment guidance, but it has not announced a release date or license.
That distinction matters: MHS is not yet a generally available open-source standard. It is a company-led research preview whose technical design and safety controls are being tested with scientific institutions, automation companies and hardware manufacturers.
1. MHS Targets the Integration Layer Between Agents and Machines
Modern laboratories and factories often contain equipment from several vendors, each with its own programming interface, data format and operating assumptions. Connecting those devices can require specialists to write separate integrations for every instrument and workflow.
Anthropic says this process commonly takes weeks or months. MHS attempts to replace much of that bespoke “glue code” with a standardized driver that translates between software and each physical device.
The driver exposes a small set of primitives, including operations such as reading a temperature or writing a new temperature setting. It also describes each device in a standard format so instruments and agents can discover one another across a network without a custom translator for every pairing.
MHS is intended to capture information that conventional APIs may omit. A driver can include natural-language tags describing properties such as the weight of a robot arm, available measurements, adjustable parameters and enforced operating limits. From those tags, MHS generates a reference file that an agent can use to understand the machine.
Operators can supply this information directly or have an agent interview them about the hardware setup. This mechanism converts knowledge that may otherwise exist only in paper manuals, local files or an experienced technician’s memory into a machine-readable description.
Once connected, an agent can control MHS equipment through three mechanisms: the Model Context Protocol, a command-line interface or code files and APIs. Code files are particularly important for long-running or time-sensitive operations because they let devices execute a deterministic sequence without waiting for the model to reason before every individual action.
Anthropic describes MHS as model-agnostic. Claude is used in the published demonstrations, but the specification is designed so other agent harnesses can access connected hardware through standard protocols.
2. Early Experiments Show Closed-Loop Hardware Control
The strongest evidence presented with the preview comes from several partner experiments. These are early, system-specific demonstrations rather than independently audited benchmarks, but they show how the common interface can support monitoring, coordination and error recovery.
At Genentech, researchers connected a liquid handler, robotic arm and microplate reader for a bicinchoninic acid protein assay. Claude initially used the same generic liquid-transfer settings for water and a viscous bovine serum albumin solution, creating bubbles and inaccurate transfers.
Researchers then asked the system to optimize the flow rates through repeated transfers and plate-reader measurements. Claude selected approximately 140 microliters per second for water, producing a root mean square error of 0.016, and 10 microliters per second for the viscous protein solution, with an error of 0.181. Genentech’s automation specialists considered both settings reasonable for that setup.
The experiment also exposed an important limitation. When bubbles caused runtime errors, Claude initially retried the operation in the same well, making the physical problem worse. Human experts had to explain that the system should use a clean well and reduce the number of mixing cycles. The resulting guidance was then encoded into reusable liquid-handling instructions.
In a Carnegie Mellon University demonstration, researchers used MHS to connect a liquid handler, plate reader, robotic arm and cameras for automated serial-dilution experiments. They reported building the device drivers and orchestration layer in about eight hours, compared with several weeks for a typical vendor-built integration.
The team deliberately introduced six unsafe or invalid conditions: a missing plate, a rotated plate, a busy reader, a disconnected camera, an unreachable device and an active emergency stop. The system blocked all six before any equipment moved.
During the subsequent experiment, the agent rejected an initial dose-response curve with an R² below 0.9, reduced the maximum concentration from 200 to 100 micrograms per milliliter and ran a second plate. That run produced an R² above 0.98 without human intervention.
A QuEra Computing pilot applied MHS to the laser system of a neutral-atom quantum computer. Four fresh Claude instances repeatedly proposed, implemented, tested and reviewed changes to a laser-recovery script. The development loop improved recovery from roughly 150 seconds with a 58% success rate to about six seconds with 96% success.
QuEra then tested the finished deterministic script without an agent controlling the laser. It successfully restored the lock in 695 of 700 randomized trials, a 99.3% success rate. The final artifact was an inspectable program rather than an AI model making every production-time adjustment.
These cases illustrate two different operating patterns. An agent can remain in the loop when conditions require continuous interpretation, or it can use live experiments to produce a deterministic procedure that is reviewed and deployed without an online model.
3. Safety Limits Are Part of the Interface, Not Proof of Safe Autonomy
Giving an AI agent access to physical equipment creates risks that do not exist when an agent is limited to documents or software. An incorrect command can damage a sample, collide robots, expose personnel to hazardous machinery or change the operating state of a sensitive instrument.
MHS addresses part of this problem by placing operating constraints in the device description. A microscope driver, for example, can enforce a laser-power ceiling independently of the agent’s requested action. The standard can also expose equipment states that must be satisfied before another device moves.
This is a more concrete control layer than asking a model to remember safety instructions in a prompt. It does not, however, establish that the complete system is safe.
Anthropic acknowledges that Claude’s physical and spatial reasoning remains limited because the model learns about the physical world largely through text and images. The Genentech foaming failure demonstrated that a model may interpret a physical problem as a software error and choose an ineffective recovery strategy.
The QuEra team similarly found that Claude could not reliably diagnose failures involving the physical rig rather than its programmatic controls. It also sometimes paused overnight while waiting for human confirmation of actions it considered risky.
MHS currently requires equipment with a programmable interface and an appropriate driver. It does not automatically make legacy machinery agent-compatible, eliminate the need for domain expertise or replace conventional safeguards such as emergency stops, access controls and physical isolation.
Anthropic says the preview will be used to build additional evaluations, best practices and a physical-safety roadmap. It plans to publish findings and deployment guidance when MHS becomes open source. Until those materials and the specification itself are public, outside developers cannot fully assess the design, enforcement boundaries or interoperability claims.
4. Hardware and Automation Partners Will Determine Whether MHS Becomes Portable
MHS originated in a collaboration between Anthropic’s Beneficial Deployments team and HHMI Janelia Research Campus. Janelia scientist Arco Bast had built a shared-memory dictionary to connect lasers, cameras and motorized focusers on a brain-imaging rig; Anthropic researcher Alek Kemeny worked with him to add AI-model access.
The preview has since expanded to hardware and automation providers. AWS plans preview support through its Strands Robots library. Automata is adding MHS support to its LINQ lab-automation platform, while Doosan Robotics is testing the standard for quality assurance and multi-robot coordination.
MBF Bioscience is developing a driver for ScanImage microscopy software. QIAGEN is testing a proof of concept with its QIAsymphony Connect nucleic-acid purification platform, and Tecan is adding support for its Fluent liquid handlers. Universal Robots has also had early access.
Hugging Face is adding MHS support to its LeRobot robotics library, while Raspberry Pi plans integrations across several products following tests of a camera driver.
For equipment vendors, the proposed advantage is the ability to describe a machine once through an MHS driver instead of supporting a separate integration for each agent framework. For laboratories and manufacturers, the potential benefit is a reusable control layer across devices that previously required independent programming.
The practical value will depend on whether drivers from different vendors behave consistently, whether safety constraints are enforced below the model layer and whether organizations outside Anthropic’s initial partner group adopt the specification. Those questions cannot be resolved while access remains application-only and the standard itself is not yet open source.
Frequently Asked Questions
Is the Model Hardware Standard available to everyone?
No. MHS is in a limited research preview, and prospective participants must apply for access.
Is MHS already open source?
No. Anthropic says it plans to open-source MHS after the research preview, but it has not provided a release date or license.
Does MHS work only with Claude?
Anthropic describes the standard as model-agnostic. Claude powered the published demonstrations, but other agent harnesses are intended to connect through protocols such as MCP.
Can MHS control any physical machine?
Not automatically. The equipment needs a programmable interface or another suitable control layer, plus an MHS driver describing its operations, state and safety limits.
Does MHS remove the need for human supervision?
No. Anthropic’s own demonstrations found failures involving physical intuition and troubleshooting that required expert guidance.
Sources
Share