Anthropic has opened a research preview of its Model Hardware Standard, or MHS, a specification intended to give AI agents a common way to operate physical equipment. The first preview is aimed at scientific laboratories and advanced manufacturers rather than general consumers.
The pitch is ambitious: instead of building a one-off software integration for every microscope, liquid handler or robotic arm, a facility could expose devices through a shared interface that an AI agent can understand. Anthropic says the work began with the Howard Hughes Medical Institute's Janelia Research Campus.
MHS is an interface layer between agents and real equipment
Anthropic says the preview can connect agents to multiple kinds of laboratory and manufacturing hardware in parallel. Its examples include microscopes, liquid handlers and robotic arms, with potential workflows ranging from routine drug-discovery experiments to laser calibration work on a quantum computer.
That does not mean Claude can simply walk into any lab and control whatever it finds. Hardware support depends on compatible integrations, permissions and the rules wrapped around each device. Ars Technica describes the project as a standardised driver-style interface: the useful idea is common plumbing, not magic universal compatibility.
The potential benefit is reduced integration work. Anthropic says labs and manufacturing facilities can currently spend weeks or months connecting specialised devices, while MHS can cut some integrations to hours or minutes. That is a vendor claim from the research preview, not an independently proven performance figure across production environments.
If the approach works broadly, the more interesting change is orchestration. An agent could coordinate several instruments, update parameters as a process runs and react to some hardware errors without a human writing a bespoke control loop for every combination of devices.
Physical control raises a different safety bar from software-only agents
Giving an AI system access to a robot arm or laboratory instrument carries consequences that are different from letting it edit a document. A bad software action can already be expensive; a bad physical action can damage equipment, waste samples or create safety risks for people nearby.
Anthropic is therefore keeping MHS in a limited research preview while it evaluates the standard with early partners. The company says safety work is part of the path toward broader availability, and the preview framing matters: this is not a finished universal industrial control standard.
That caution also affects how the integration-time claim should be read. Fast setup in a controlled trial does not remove the need for site-specific validation, access controls, emergency stops, calibration procedures or the existing safety systems around industrial and scientific equipment.
The collaboration with Janelia also gives the preview a concrete scientific setting rather than treating physical agents as a purely hypothetical demo. The harder question is whether the same interface can stay dependable when equipment, procedures and risk controls vary widely between facilities.
For now, MHS is best understood as an attempt to standardise the connection point between AI agents and hardware. The immediate milestone is the research preview with selected labs and manufacturers. The next meaningful evidence will be how well those integrations hold up across more devices and whether Anthropic can demonstrate safety controls strong enough for broader deployment.
Reporting notes