---
title: MHS (Model Hardware Standard)
type: vocabulary
url: "https://www.envisioning.com/vocab/mhs-model-hardware-standard"
summary: "Anthropic's shared specification that lets any AI agent safely operate physical lab and manufacturing instruments via standardized drivers."
year: 2026
generality: 0.55
---

# MHS (Model Hardware Standard)

Anthropic's shared specification that lets any AI agent safely operate physical lab and manufacturing instruments via standardized drivers.
Title: Model Hardware Standard (MHS)
Slug: model-hardware-standard

The Model Hardware Standard (MHS) is a shared specification introduced by Anthropic in August 2026 in collaboration with HHMI Janelia Research Campus. It allows any AI agent, not just Claude, to safely operate physical laboratory and manufacturing devices. MHS is to hardware what the Model Context Protocol (MCP) is to software tools: a standardized interface layer that lets heterogeneous instruments be discovered, controlled, and orchestrated by an agent without bespoke integration work for each device. The specification was motivated by the observation that lab automation has historically required weeks or months of specialist work to integrate per-instrument drivers, leaving most academic and even industrial laboratories with islands of incompatible equipment that an AI cannot span.

The technical architecture has three layers. At the bottom, an MHS driver is a piece of software that translates between a device's native programmable interface and a small set of universal primitives, including read (e.g., "get temperature") and write (e.g., "set temperature"). Each driver exposes a standardized manifest describing the device's general characteristics: what it can measure, what can be adjusted, and what safety limits are enforced. The manifest makes the device discoverable across networks and immediately accessible to any agent that speaks the protocol. The manifest also supports natural-language tags, so information traditionally locked in paper manuals or tacit knowledge (the weight of a robotic arm, the maximum flow rate before bubbles form) can be written into the driver file directly, either by a human or by an agent that interviews the user about their setup. Above the driver layer, three control mechanisms are supported: MCP for fine-grained agentic control, a command line interface for scripted operation, and code files containing chained driver commands for long-running tasks where the agent would otherwise need to reason at every step.

The first wave of partner deployments illustrates the practical span of the standard. Genentech used MHS to automate a BCA protein assay across a liquid handler, a robotic arm, and a microplate reader, with Claude orchestrating the protocol and recovering from hardware errors mid-run. The University of Washington Baker and Pinglay labs used MHS to remote-monitor instruments, supervise a qPCR run that halts the reaction at the right amplification moment, and coordinate plate handoffs between robotic arms and liquid handlers without collisions. Other partners span biotech, robotics, electronics, and quantum computing, including laser calibration on a quantum computer. The standard is designed to be model-agnostic, so any agent harness that speaks MCP can use MHS without modification.

MHS is adjacent to but distinct from several existing concepts. The Model Context Protocol (MCP) standardizes how AI agents connect to software tools and data sources; MHS uses MCP as one of its three control mechanisms but adds the standardized hardware-driver layer underneath. Computer use refers to AI agents operating software GUIs; MHS refers to AI agents operating physical instruments through programmatic interfaces. Self-driving laboratories apply closed-loop AI control to experiments; MHS is the protocol that makes such experiments cross-vendor and cross-instrument. Lab automation as a field predates the AI era and includes mechanical and analog systems; MHS is specifically the AI-native protocol layer that makes those systems programmatically accessible to agents. Anthropic has announced that MHS will be open-sourced after the research preview, with the explicit goal of becoming a community standard analogous to MCP.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/mhs-model-hardware-standard)