What MHS Is (and Isn't)
The hardest part of lab automation is rarely the experiment itself — it's getting instruments from different vendors to talk to each other. Integration typically takes weeks to months. On August 27, Anthropic announced the Model Hardware Standard (MHS): a shared specification that lets AI agents safely operate physical devices, aiming to compress integration from weeks/months to hours/minutes.
MHS is a pure software specification. It defines no voltages, connectors, or mechanical dimensions. It has four layers:
| Module | Purpose | Example | |---|---|---| | Standard drivers | Translate hardware into two primitives | read temperature, set temperature | | Device discovery | Devices self-describe in a standard format | agents find devices on the network automatically | | Metadata tags | Natural-language device characteristics, auto-generated reference docs | payload limits, adjustable ranges, safety limits | | Three control channels | MCP, command line, direct API calls | orchestrate a chain of devices in one line of code |
Safety limits are mandatory and enforced at the driver layer. MHS is model-agnostic — Claude can use it, and so can other vendors' agents.
The stack originated from an unconventional source: Arco Bast, a postdoc at Janelia Research Campus, wrote a shared-memory dictionary for his brain-imaging rig; Anthropic's Alek Kemeny saw it, and the two grew it into a standard.
Six Pilots, Hard Numbers
Research preview participants: Genentech, University of Washington, Carnegie Mellon, HHMI Janelia, QuEra, and Tetsuwan Scientific.
Carnegie Mellon — serial dilution. A CyBio Felix liquid handler, a Varioskan LUX plate reader (no API — the agent drove the GUI directly), a Spinnaker arm, and cameras spread across three incompatible computers were integrated in ~8 hours versus the usual weeks. The agent caught all six deliberately injected faults (missing plate, misaligned plate, occupied reader, camera disconnect, device loss, e-stop). When one run's R² fell below 0.9, the agent adjusted concentrations and re-ran autonomously, reaching R² above 0.98.
QuEra — laser relocking. For neutral-atom quantum computers, the legacy script took 150 seconds to recover with 58% success; human experts took 5–10 minutes. The agent-written relock controller succeeded 695/700 times with recovery in seconds. Across 363 experiments and 16 hours unattended, residual phase error dropped from 15.7 mV to 1.55 mV; tuning was never lost over 19 hours, versus ~1.6 losses/hour for a human expert.
Genentech — BCA protein quantification. Claude orchestrated three devices and self-optimized pipetting speeds (140 µL/s for water, 10 µL/s for viscous BSA), recovering from tip-pickup and liquid-level errors. The only human intervention needed: distinguishing bubble-related physical issues from software issues.
Other results: University of Washington connected six instruments within a week for protein design; Janelia unified seven vendors' software into one state dictionary for zebrafish imaging; Tetsuwan's 9,143th dispensing beat vendor parameters by ~12%.
Ecosystem
Vendors integrating: AWS, Danaher, Tecan, QIAGEN, Doosan, Universal Robots, Automata, MBF Bioscience. Early adopters include Hugging Face's LeRobot and Raspberry Pi (official MHS camera driver) — meaning a $100 single-board computer can be safely agent-driven.
Why This Matters Beyond Lab Automation
While embodied AI focuses on giving models bodies (humanoids, end-to-end VLA, world models), MHS standardizes the agent-to-device interface instead — analogous to USB for peripherals or MCP for tool calling. If open-sourced and widely built in, agent access to labs shifts from bespoke projects to installing a driver. Impact on science is direct: University of Washington notes a de novo protein design round costs ~$100 and a week to validate while the design itself costs one cent — the bottleneck is entirely wet-lab side.
Caveats
Anthropic itself acknowledges:
- Claude's spatial/physical reasoning still needs expert oversight; QuEra's agent was sometimes overcautious, stopping overnight to await approval;
- Compute costs for long-running agents are nontrivial;
- Orchestration of complex protocols still requires tuning.
- Anthropic announcement: https://www.anthropic.com/news/model-hardware-standard-research-preview
- Application portal: https://modelhardwarestandard.com
Additionally, MHS requires a programmable interface (even a GUI qualifies); purely mechanical devices are out of scope for now.
Outlook
The key detail to watch is the open-source commitment: closed standards don't generate network effects. Anthropic's ambition is to define the device protocol layer of the agent era — an "MCP for the physical world." The signal to watch within six months: whether MHS compatibility becomes a marketing bullet in Tecan/Danaher product catalogs.
References: