We’re opening a investigation preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely run beingness devices, to a first group of technological investigation labs and precocious manufacturers. MHS enables AI agents to run aggregate laboratory and manufacturing instruments, specified arsenic microscopes, liquid handlers, and robotic arms, successful parallel, and execute intricate tasks ranging from regular supplier find experiments to laser calibration connected a quantum computer. The improvement of MHS began arsenic a collaboration betwixt Anthropic and HHMI Janelia Research Campus.
It typically takes a laboratory aliases manufacturing installation weeks, if not months, to group up and merge their hardware. Most devices don’t pass pinch each other, alternatively requiring specialists to build bespoke integrations. MHS reduces this integration activity to hours aliases minutes. And by incorporating AI into these tools, MHS besides helps researchers and engineers much readily orchestrate autonomous, round-the-clock experiments and workflows, pinch agents capable to logic done each measurement successful an experiment, update parameters successful existent time, and, successful immoderate cases, retrieve from hardware errors without intervention.
We’re sharing an early type of MHS pinch partners crossed science, robotics, electronics, and manufacturing truthful we tin collaborate to build information evaluations and create champion practices for AI systems operating beingness equipment, up of making the modular unfastened source. MHS useful pinch immoderate instrumentality that has a programmable interface. It is besides model-agnostic, and immoderate supplier harness tin entree it utilizing modular protocols, specified arsenic the Model Context Protocol. To use for entree to the investigation preview, head here.
How MHS works
Getting aggregate devices successful a laboratory aliases connected a mill level to pass pinch 1 different tin beryllium challenging, moreover mounting speech the added trouble of integrating AI into the setup. Each instrumentality tends to person its ain programming interface, and truthful acold location has been nary standardized measurement to merge them. And erstwhile the devices are connected, location is nary communal measurement for them to stock information pinch an AI agent, nor to fto the supplier run them safely.
MHS addresses these challenges by introducing a standardized driver: package that translates betwixt a computer’s operating strategy and a hardware device. The MHS driver uses a elemental group of primitives—commands for illustration “read” (for example, “get temperature”) aliases “write” (for example, “set temperature”)—that immoderate hardware instrumentality tin understand and enactment on. And it makes each instrumentality discoverable successful a modular format, truthful that devices and agents tin find each different and pass crossed networks without needing a bespoke “translator” programme successful between.
The MHS driver besides helps an AI supplier understand really to usage a instrumentality it has ne'er seen before, giving it accusation astir instrumentality characteristics that whitethorn not beryllium discernable from codification unsocial (for example, the weight of a robot arm, which is important for knowing really to manipulate it safely). To date, overmuch of this accusation has been stored successful insubstantial manuals, connected a user’s computer, aliases arsenic tacit knowledge. But the MHS driver contains tags that fto the personification constitute this accusation straight successful earthy connection (users tin either do this themselves, aliases by chatting to an supplier that interviews them astir their hardware setup). With the accusation from these tags, the MHS driver past automatically produces a reference record pinch accusation astir a device’s wide characteristics, specified arsenic what it tin measure, what tin beryllium adjusted, and what information limits will beryllium enforced. This record gives the supplier everything it needs to cognize to run the device.
After the devices are connected and the supplier knows really to usage each one, the supplier needs a measurement to power the hardware. For MHS, location are 3 specified mechanisms: MCP, the bid statement interface, and codification files (APIs). These activity together to alteration orchestration crossed aggregate devices via a azygous statement of code.
Once the supplier tin power the devices, it’s capable to person operating information from each 1 and supervise and nonstop the activity astatine a precocious level. The supplier tin series steps crossed instruments, show results, and set parameters arsenic conditions alteration successful existent time. When the supplier needs to execute long-running tasks aliases run devices faster than its online reasoning would allow, it tin concatenation together driver commands from 1 aliases much devices successful codification files. This allows the devices to transportation retired operations themselves, without the supplier needing to logic astatine each step.
As we’ve tested MHS, we’ve recovered that Claude interacts pinch experiments and hardware successful an exploratory manner, overmuch arsenic a intelligence would. For example, we observed Claude make an accommodation to a laser, observe the results done a camera to measure really its accommodation moved the laser beam, and repetition the process, seeking to understand the series of events. Claude past packaged what it learned into codification files, penning a deterministic book that fto it align the laser without having to logic astatine each step, truthful the full process could tally arsenic a azygous command.
Early examples from MHS
We are only conscionable opening to spot what group tin do pinch frontier models and MHS, but our dream is that the modular tin beryllium of usage to researchers, engineers, and different practitioners successful speeding up the process of find and experimentation successful immoderate domain that uses devices pinch a programmable interface.
As we developed MHS, we shared it pinch a fistful of labs and hardware manufacturers successful biotech, robotics, quantum computing, and different fields. Across these early projects, we saw MHS trim the clip it took to merge devices, make it imaginable to iterate faster successful a assortment of experimental settings, and assistance pinch the unrecorded cognition of machines and real-time responsibility detection. Below, our partners stock the specifications of immoderate of their early projects involving MHS.
Genentech: Implementing MHS for laboratory automation
Researchers astatine Genentech implemented and tested MHS arsenic a impervious of conception for automating the BCA macromolecule assay, a modular process to measurement full macromolecule attraction successful a sample, which requires coordinating crossed a liquid handler, a robotic arm, and a sheet reader.
University of Washington Baker and Pinglay labs: Bringing AI agents to the bench
Zihao Song, a PhD student successful the University of Washington Baker and Pinglay labs, utilized MHS to build a dashboard to remotely show his instruments; an AI agent-supervised qPCR (which copies a target DNA series done repeated cycles of heating and cooling) that watches amplification curves and halts the process astatine the correct moment; and an integration betwixt a robotic limb and a liquid handler for collision-free sheet handoffs.
Carnegie Mellon University: Determining dose-response curves done accelerated automation
Researchers astatine Carnegie Mellon University utilized MHS to tally serial dilution dose-response experiments astir 3 times faster than before, pinch an AI supplier orchestrating a liquid handler, a sheet reader, a robotic arm, and monitoring cameras dispersed crossed 3 computers pinch fundamentally incompatible interfaces.
HHMI Janelia: Using MHS to accelerate microscopy research
At HHMI Janelia Research Campus, researchers are utilizing MHS to velocity up a scope of microscopy-related projects. Here, Virginie Ruetten, a intelligence successful the Ahrens laboratory who studies really slumber helps the assemblage retrieve from stress, shares really she utilized MHS to unify and orchestrate a rig that antecedently progressive 7 different vendor programs without a shared interface.
QuEra Computing: Using MHS successful quantum laser stabilization
QuEra, a institution that builds quantum computers utilizing neutral atoms, utilized MHS to springiness an AI supplier power complete parts of the laser strategy wrong its quantum machines. The supplier developed a controller that recovers the laser’s “lock”—the ultra-precise wave the lasers must clasp to interact pinch the atoms—99.3% of the clip without quality intervention.
Tetsuwan Scientific: Using MHS to tally qPCRs to floor plan section contamination
Researchers astatine Tetsuwan integrated MHS pinch its automated biology laboratory platform, ResearchOS. MHS helped orchestrate a qPCR workflow to lend to national subject efforts to qualify contamination successful California’s San Pedro Creek.
Hardware vendors and the package companies that support them are besides building MHS support into their instrumentality truthful AI agents tin much easy observe and run their devices. For example:
- Amazon Web Services will support MHS done Strands Robots, the room for connecting AI agents to beingness devices. AWS will supply participants a private, pre-release type of the Strands Robots package for the long of the MHS investigation preview.
- Automata is adding MHS support to LINQ, their laboratory automation platform, to execute intelligent correction handling of instruments successful autonomous labs.
- Danaher and Anthropic are actively exploring really MHS-supported capabilities could alteration its smart instruments and autonomous laboratories to standard biomedical investigation and development.
- Doosan Robotics is testing MHS pinch their robotic arms, including to execute automated value assurance and coordinate tasks crossed aggregate robots.
- MBF Bioscience is building an MHS driver for ScanImage, the package that runs laser-scanning microscopes successful hundreds of neuroscience labs worldwide, to merge AI agents into real-time information study and experiments.
- QIAGEN is experimenting pinch MHS done a moving impervious of conception connected its nucleic acerb purification platform, QIAsymphony Connect, showing really AI agents could thief laboratories troubleshoot instrumentality issues faster, guideline operators done recovery, and amended instrumentality uptime while reducing consequence to biologic samples.
- Tecan is adding MHS support for their Fluent liquid handling platforms truthful AI agents tin observe and run them directly.
- Universal Robots has had early entree to MHS and plans to adhd support to its robotics platform.
Joining the investigation preview
These early results from our partners are encouraging, but we person much activity to do connected the modular earlier we open-source it. As a ample connection model, Claude learns astir the beingness world done matter and images, meaning its spatial and beingness reasoning person limitations that still require master oversight. When moving pinch macromolecule samples, for example, Genentech researchers had to guideline Claude to admit that errors caused by foaming successful samples were beingness failures, not package bugs, that could only beryllium mitigated done the due beingness corrections.
MHS besides doesn’t yet activity pinch hardware that lacks a programming interface, truthful we’re moving pinch the manufacturers of specified devices to build successful MHS drivers. Many developers already usage Claude Code to activity pinch individual pieces of beingness equipment; for the adjacent shape of MHS, we dream to grow the modular to screen much of the devices developers build on. Early adopters see Hugging Face, who are adding MHS support successful LeRobot, their robotics library, and Raspberry Pi, who are enabling MHS integration crossed a number of their products pursuing successful tests utilizing their Camera MHS Driver.
We will besides usage the investigation preview to build further information evaluations pinch our motorboat partners and fortify protections for the usage of AI successful the beingness world. We are processing a beingness information roadmap to further bolster our safeguards argumentation and enforcement sum against the consequence of misuse. When we open-source MHS, we will merchandise findings from the investigation preview arsenic portion of our guidance for deploying the modular safely.
We’re inviting stakeholders crossed industries to subordinate the waitlist for our investigation preview of MHS. If you’d for illustration to participate, submit your liking here.
Acknowledgments
MHS began arsenic a collaboration betwixt Alek Kemeny connected Anthropic’s Beneficial Deployments team and Arco Bast, a postdoctoral intelligence astatine HHMI Janelia Research Campus. Bast was moving analyzable brain-imaging experiments connected a rig that mixed lasers, motorized focusers, and specialized cameras from different vendors pinch nary communal interface. To velocity up his experiments, he developed a shared representation dictionary that enabled the instruments to pass pinch 1 different astatine representation speed. Kemeny and Bast worked together to merge AI models into that interface.
We convey everyone who has contributed to this activity truthful far, including, but not constricted to, Aaron Boswell, Ben Arthur, Boaz Mohar, Gagan Bhat, Mark Kittisopikul, Nadine Yasser, Nick Purcell, Takashi Kawase, and Virginie Ruetten. We look guardant to moving MHS guardant pinch our manufacture partners and, soon, pinch the open-source community.
English (US) · 
Indonesian (ID) · 
