How AI Is Gaining Lab Hands and Eyes: Inside Anthropic’s Model Hardware Standard

Conclusion: MHS standardizes device semantics, not robots

MHS targets any device with a programmable interface, without tying to a specific model or agent framework; agents can access devices via MCP, command‑line interfaces, or code/API files. The standard extends the host‑client‑server architecture already used by software tools to the physical layer.

How MHS Works

Laboratory and factory devices normally have disparate interfaces, drivers, vendor software, data formats, and operating habits, making integration weeks or months long. MHS introduces a standardized driver that abstracts the OS‑hardware relationship into simple primitives such as read (e.g., temperature) and write (e.g., set temperature). Each device is discovered in a uniform format, enabling cross‑network communication with agents without pairwise translators.

Beyond commands, a driver also records device semantics—weight, measurable parameters, adjustable settings, and safety limits—often hidden in manuals or expert knowledge. Users can add natural‑language tags; agents can interview users to capture these facts, after which the driver auto‑generates a reference file for the agent.

After devices are discovered and understood, the control layer offers three paths: MCP , CLI , and code/API files . MCP and CLI suit interactive orchestration, while code/API files allow deterministic scripts to chain driver commands, letting agents schedule instrument sequences, monitor results, and adjust parameters on the fly, while long‑running or low‑latency tasks run offline.

Six Early Case Studies

1. Genentech – Closed‑loop BCA protein assay

Using MHS, a liquid‑handling workstation transferred liquids, a robotic arm moved consumables, and a plate reader measured absorbance. Claude coordinated the protocol and hardware communication. The closed‑loop test achieved flow rates of 140 µL/s (RMSE 0.016) for water and 10 µL/s (RMSE 0.181) for viscous BSA, matching expert performance.

When bubbles formed, Claude initially retried the same well, worsening the problem. Researchers taught Claude that bubbles are a physical failure, not a software bug, prompting a change to a clean well and reduced mixing. The experience was encoded as a reusable pipetting skill for different fluid properties.

2. University of Washington – qPCR monitoring and plate hand‑off

Designing a protein from scratch costs as little as $0.01 computationally, but testing a candidate on a bench costs $100 and a week of labor , often for ~1,000 candidates. Traditional automation struggles with frequent protocol changes and multi‑vendor integration.

MHS gave the agent “eyes, hands, and sense of time”: it read instrument states, ran devices, and coordinated them. The agent streamed real‑time qPCR curves, asked the researcher whether to stop, and upon receiving a stop command applied a 4 °C hold to prevent DNA degradation.

In a plate‑hand‑off demo, a liquid‑handler dispensed, a LeRobot arm moved the finished plate, and Claude Code triggered the next step only after the previous one completed. Six instruments were integrated in under a week, though the authors note that complex protocols still need extensive optimization and compute cost trade‑offs.

3. Carnegie Mellon University – Autonomous dilution‑response run

A three‑computer setup (liquid processor, plate reader, robotic arm, monitoring camera) performed serial dilutions and readouts. MHS compressed the workflow to 3× faster . The system injected six failure modes (missing plate, plate rotation, busy reader, camera disconnect, device unreachable, emergency stop) and blocked them before any device moved.

The first run saturated at high concentration (R² < 0.9); the agent discarded the plate and reduced the top concentration from 200 µg/mL to 100 µg/mL . The second run achieved R² > 0.98 with 3.4 % variance, fully autonomous. Total time from raw setup to a self‑rerun curve was about 8 hours , compared with weeks for vendor solutions.

4. HHMI Janelia – Consolidating seven imaging programs

Researchers used two‑photon microscopes, lasers, detectors, and four translation stages from different vendors, each controlled by separate software (MATLAB, Python, C#). Starting an experiment required launching seven programs in a fixed order.

MHS replaced point‑to‑point connections with a shared in‑memory state dictionary. Adding a new camera now takes minutes instead of days, and experiment startup collapsed to a single dashboard click.

The system also unified data streams (heart rate, images, time series, spectra) into a documented format, enabling online heartbeat tracking and predictive feedback to downstream processes.

5. QuEra – Auditable laser‑lock scripts for neutral‑atom quantum computers

Laser frequency stability of 1 × 10⁻¹² is required; temperature, vibration, and pressure can cause lock loss. Human‑written recovery scripts had a 58 % success rate and took ~150 s each.

With MHS, four Claude instances generated, executed, and logged scripts autonomously overnight, reducing recovery time to 6 s with a 96 % success rate. In 700 random perturbation tests, the system relocked 695 times (99.3 %) . The final script is deterministic, auditable, and runs without continuous agent control.

Claude also tuned twelve interrelated PID parameters by collecting oscilloscope traces, computing full spectra, and iteratively minimizing residual noise.

QuEra notes that hardware faults still expose the model’s limited physical understanding; risky actions often pause for human approval.

6. Tetsuwan – Natural‑language protocols for environmental‑sample qPCR

ResearchOS translates natural‑language protocols into a custom experimental language, then compiles to automation code. Using MHS, the team linked a liquid‑handling robot, a robotic arm, and consumables to run a citizen‑science project on Pacifica’s San Pedro Creek pollution.

Viscosity‑induced bubbles were detected by a camera; the system suggested centrifugation to settle bubbles, converting the instruction into device‑specific parameters via the driver.

Across 9,143 independent dispenses, 300 transfer types, and 1,508 measurement conditions, Claude‑MHS predictions improved pipetting precision by 12 % (sign test p≈0.001) and increased repeat‑run performance by ~17 %.

qPCR of environmental samples identified E. coli and Bacteroides markers, confirming human fecal contamination as the primary source; the authors caution against directly comparing mean Ct values across different markers.

Early Adopters and Ecosystem

Listed early supporters include AWS (Strands Robots), Automata (LINQ), Danaher, Doosan Robotics, MBF Bioscience (ScanImage driver), QIAGEN (QIAsymphony Connect PoC), Tecan (Fluent liquid‑handling), Universal Robots, Hugging Face (LeRobot support), and Raspberry Pi (camera driver).

Author’s Perspective

The most significant insight is that the preview goes beyond “agents moving a button.” In the QuEra case, the model’s exploration is distilled into deterministic, auditable scripts; Genentech captures bubble‑handling expertise as reusable skills; Janelia creates a universal state dictionary for cross‑device data.

However, a unified interface does not automatically grant physical understanding. Models may misinterpret bubbles as software bugs, pause for human approval on risky actions, and cannot control devices lacking programmable interfaces. Anthropic plans further safety evaluations and will publish deployment guidelines when the standard is open‑sourced.

Overall Assessment

MHS functions as a “device semantic layer” and orchestration tier for laboratory automation rather than a universal controller that replaces scientists. It can free researchers from repetitive monitoring and manual integration, but hardware interlocks, safety authorizations, material‑level judgments, and on‑site supervision must still be explicitly designed into systems.

Sources: [1] https://www.anthropic.com/news/model-hardware-standard-research-preview; [2] https://modelcontextprotocol.io/docs/learn/architecture; [3] https://www.modelhardwarestandard.com