MHS: Claude Can Now Operate Lab Robots and Physical Devices
On August 27, 2026, Anthropic opened a research preview of the Model Hardware Standard (MHS) — a shared specification that lets AI agents control microscopes, robotic arms, liquid handlers, and other instruments in parallel. MHS reduces device integration from weeks to hours. This lesson shows how Claude behaves in physical lab settings and what you need to supply to work with it effectively.
Claude can now operate physical lab equipment. The Model Hardware Standard (MHS) is a shared specification that lets AI agents control instruments like microscopes, liquid handlers, robotic arms, and quantum computing hardware — simultaneously, around the clock.
Before MHS, connecting multiple lab devices required specialists to build custom integrations for each pair of instruments. That process typically took weeks or months. MHS cuts it to hours or minutes.
MHS began as a collaboration between Anthropic and HHMI Janelia Research Campus. It works with any device that has a programmable interface. It is model-agnostic — any agent harness can access it using standard protocols, including the Model Context Protocol (MCP).
MHS works by giving every device a standardized driver. A driver is software that translates between a computer and a hardware device. The MHS driver uses simple commands — "read" (for example, "get temperature") and "write" (for example, "set temperature") — that any device can understand.
Each device also becomes discoverable across a network in a standard format. Devices and agents find each other without a custom translator in between.
The driver includes tags where users write physical facts in natural language — the weight of a robot arm, its safety limits, what it can measure. Users can fill these tags themselves, or by chatting with Claude, which interviews them about their hardware. The driver then produces a reference file so Claude knows how to operate the device safely, even one it has never seen before.
Once devices are connected, Claude controls them via MCP, the command line, or code files. It sequences steps across instruments, monitors results, and adjusts parameters in real time.
Claude approaches experiments the way a scientist does — iteratively. In testing MHS, Anthropic observed Claude adjusting a laser, checking the result through a camera, adjusting again, and repeating until it understood the sequence. Once Claude grasped the pattern, it packaged its learning into a code file — a deterministic script that aligned the laser as a single command, without reasoning at every step.
When Claude needs to run long tasks or operate devices faster than real-time reasoning allows, it chains driver commands into code files. The devices then execute those operations independently.
Claude can also recover from hardware errors without human intervention — a capability most current scientific instruments lack on their own.
Genentech used MHS to automate the BCA protein assay — a three-instrument procedure. The BCA (bicinchoninic acid) assay measures total protein concentration in a sample. It requires coordinating a liquid handler, a robotic arm, and a plate reader.
Genentech asked Claude to optimize liquid transfer flow rates for two fluids: plain water and viscous bovine serum albumin (BSA). Claude ran trial transfers, analyzed plate reader data, and minimized the gap between its results and an expert's reference transfer. It arrived at approximately 140 µL/s for water (0.016 RMSE) and 10 µL/s for BSA (0.181 RMSE) — values Genentech's automation experts confirmed as reasonable.
Claude also recovered on its own from tip pickup failures and fluid detection errors. But there was a clear limit: when bubbles formed in the liquid, Claude's instinct was to retry in the same well with different parameters. That made the problem worse. Claude lacked physical intuition about bubble physics. Once Genentech told Claude that bubbles required moving to a clean well and reducing mixing cycles, Claude held that context for the rest of the run. Genentech then codified those steps as reusable liquid-handling skills.
The practical takeaway: Claude reasons powerfully about sequences and data. Physical domain knowledge — especially for error conditions — still needs to come from you.
What MHS means for working with Claude:
MHS lets Claude control lab and manufacturing hardware through a standardized driver, cutting device integration from weeks to hours.
Claude sequences steps across instruments, adjusts in real time, and can recover from hardware errors autonomously.
Physical facts — weight, safety limits, fluid behavior — must be supplied in MHS driver tags or in your prompts. Claude cannot infer them from code alone.
When Claude's error recovery instinct conflicts with physical reality (like retrying in a foamy well), give it the physical context it's missing rather than asking it to retry the same approach.
Package what Claude learns into reusable code files or skills so that knowledge carries forward to future runs.
MHS is in research preview as of August 27, 2026, and will be open-sourced after safety evaluations and best practices are established with partners.