Anthropic today opened a research preview of its Model Hardware Standard (MHS), a shared specification that lets AI agents safely operate physical instruments in labs and factories. Partner labs report slicing device integration time from weeks or months to hours or minutes, the clearest signal yet that agentic AI is moving from screens into the physical world.
What is the Model Hardware Standard?
MHS, developed with HHMI Janelia Research Campus, is a protocol layer for physical devices. Its core is a standardized driver: software that translates between an agent and a machine, built on simple read/write primitives such as "get temperature" or "set temperature" that any device can understand. Every device becomes discoverable in a standard format, so agents and instruments can find each other across a network without bespoke translator programs in between.
The driver also carries natural-language tags describing physical characteristics a model cannot infer from code alone, such as the weight of a robotic arm or the safety limits to enforce. From those tags it generates a reference file, listing what a device measures, what can be adjusted, and what is off limits, so an agent understands an unfamiliar instrument before touching it.
Control happens through three mechanisms that work together: the Model Context Protocol (MCP), a command line interface, and code files (APIs). Together they let a single line of code orchestrate multiple devices. For long-running operations, agents chain driver commands into deterministic scripts, so instruments execute autonomously without the agent reasoning at every step.
Why it matters: MCP for the physical world
The pattern is familiar: MCP gave agents a standard way to read software context and call tools. MHS extends the same idea to hardware, replacing vendor-specific SDKs with one driver model any agent harness can use. Because MHS is model-agnostic and reachable over MCP, it slots into existing agent stacks instead of demanding new infrastructure.
The economic effect is an order-of-magnitude drop in integration cost. Labs that once spent weeks wiring a liquid handler, a robotic arm, and a plate reader together can stand up the same rig in hours. The bottleneck shifts from connecting machines to deciding which experiments to run, which is exactly where frontier models are strong. It is worth reading this alongside the race for shared robotic brains: as embodied AI models consolidate, the hardware interface layer is consolidating too.
Early proof points from partner labs
The preview already spans biotech, robotics, and quantum computing:
- Genentech automated a BCA protein assay, coordinating a liquid handler, a robotic arm, and a plate reader as a proof of concept.
- University of Washington researchers built a dashboard for remote instrument monitoring, an AI-supervised qPCR that halts amplification at the right moment, and collision-free plate handoffs between arm and handler.
- Carnegie Mellon ran dose-response experiments about three times faster, orchestrating instruments across three computers with fundamentally incompatible interfaces.
- HHMI Janelia unified a microscopy rig that previously required seven vendor programs with no shared interface.
- QuEra Computing gave an agent control over parts of the laser system inside its quantum machines; the agent recovers the laser lock 99.3 percent of the time without human intervention.
- Tetsuwan Scientific used MHS to run qPCR workflows for a citizen-science survey of pollution in California's San Pedro Creek.
Hardware vendors are building MHS in as well: AWS through its Strands Robots library, Universal Robots, Doosan, Tecan's Fluent liquid handlers, QIAGEN's QIAsymphony Connect, Danaher, MBF Bioscience's ScanImage, and Automata's LINQ platform. Hugging Face plans MHS support in its LeRobot robotics library, and Raspberry Pi is enabling integration across its camera products.
What it means for the industry
The strategic read: instruments become endpoints on the agent network. When a vendor ships MHS support, its devices are agent-discoverable and agent-operable out of the box, the same move MCP made for enterprise software and the AI-Native SDLC playbook made for engineering workflows. Autonomous round-the-clock experiments become a staffing decision rather than an engineering project.
Safety is the deliberate gate. Anthropic is using the preview to build physical safety evaluations with launch partners and a physical safety roadmap, and it acknowledges the limits: Claude reasons about the physical world through text and images, so expert oversight still matters. Genentech, for example, had to teach the model that foaming in protein samples was a physical failure, not a software bug. MHS enforces per-device safety limits, and findings will be released when the standard is open sourced.
For agent builders, durable agent state is the next frontier: agents that supervise experiments for hours need to resume, replay, and reconcile long-running work, capabilities worth studying in any serious agent stack.
What you can do now
- Apply for the research preview at modelhardwarestandard.com if you run lab or manufacturing hardware.
- Ask instrument vendors whether MHS support is on their roadmap; AWS, Universal Robots, Doosan, and Tecan already ship or plan it.
- Developers: treat physical devices as MCP endpoints and watch Strands Robots and LeRobot for agent-hardware integrations.
- Hold production workflows until the open-source release and its safety evaluations land.
FAQ
Q: What is the Model Hardware Standard (MHS)?
A: A shared specification from Anthropic and HHMI Janelia that lets AI agents operate physical devices through a standardized driver. It cuts integration time from weeks or months to hours or minutes and works with any device with a programmable interface.
Q: How does MHS differ from MCP?
A: MCP standardizes how agents access software data and tools; MHS standardizes how agents control physical hardware, including device discovery and enforced safety limits. MHS is accessible over MCP, so the two compose rather than compete.
Q: When will MHS be open sourced?
A: Anthropic will open source the standard after the research preview, alongside findings from partner deployments and a physical safety roadmap. Early results include QuEra recovering its quantum laser lock 99.3 percent of the time without human intervention.
