Anthropic Model Hardware Standard (MHS)

MHS standardizes physical-hardware integration through a driver layer that translates between a computer's operating system and a device using simple primitive commands - read and write - plus discovery, natural-language device tags, and reference files that spell out safety limits and operational parameters before an agent ever touches the equipment ^[1]. That structure lets an agent enumerate what is connected and what it is allowed to do first, then act. Once a device is described this way, agents can drive it through any of three paths - the Model Context Protocol, a command-line interface, or direct code and API calls - and because the standard itself is model-agnostic, any LLM harness can use it, not only Claude ^[2]. Anthropic is treating this as the physical-world counterpart to its 2024 Model Context Protocol launch: the company's Jonah Cool has said research often stalls not for lack of ideas but because the equipment can't be used, and technical staff member Alek Kemeny has framed MHS as kind of like the USB for AI to software connection ^[4]- the same unifying pitch MCP made for software, now aimed at lab instruments, robotic arms, and manufacturing hardware that previously required specialists to write bespoke, incompatible integrations for each vendor ^[1].

The strongest case for MHS is quantitative. Carnegie Mellon orchestrated a liquid handler, plate reader, robotic arm, and monitoring cameras across three computers with previously incompatible interfaces, taking a serial dilution dose-response experiment from driver-writing to a finished curve - including an autonomous rerun after the agent rejected a low-R^2 fit - in about eight hours, versus several weeks with a standard vendor setup ^[1]. QuEra Computing's numbers are even sharper: an MHS-driven agent building a deterministic laser-locking controller raised relock success from a 58% custom-script baseline to 99.3% across 695 of 700 trials, cut per-attempt time from roughly 150 seconds to 10-14 seconds, reduced servo residual error from 15.7 mV to 1.55 mV, and held a lock steady for 19 hours without slipping ^[1]. Genentech automated a BCA protein assay and had Claude tune liquid-handling flow rates for water and viscous BSA solutions to sub-0.2 RMSE accuracy, while HHMI's Janelia Research Campus - an MHS development partner - unified seven vendor programs controlling lasers, mirrors, and detectors for two-photon zebrafish imaging, cutting new-camera integration from days to minutes ^[3]. Anthropic frames the entire preview around that compression, describing MHS as reducing integration work to hours or minutes ^[5].

Not everyone is convinced MHS is a leap rather than a rebrand. In community discussion of the launch, a recurring objection is that MHS is just another protocol layered on top of MCP and existing APIs - agents, some argued, could already interface with hardware specs through Claude Code, making the marginal value of a new standard unclear. A more technical strand of pushback held that MHS is poorly suited to low-latency, deterministic embedded control, such as PID-loop processes, because agentic queuing is not designed for the speed those systems need; others countered that MHS targets multi-step bio-lab and research-device chains rather than single real-time control loops, a narrower and more defensible claim. The partner results give that narrower claim some support - the CMU, QuEra, Genentech, and Janelia cases are all multi-step orchestration tasks, not millisecond-scale control - but they also expose the limits of AI physical reasoning: because large language models learn about the physical world from text and images rather than direct sensation, Genentech's researchers had to explicitly teach their agent that sample foaming was a hardware fault rather than a software bug, a reminder that expert oversight remains necessary even where MHS works as advertised ^[6].

MHS is arriving well ahead of any regulatory framework built for it. The European Union's Machinery Regulation 2023/1230, which takes effect January 20, 2027, will for the first time cover AI-based safety functions and self-evolving machinery, and could require notified-body approval for higher-risk categories of equipment that agents operate autonomously - yet no UK regulator has staked out a clear position on AI agents running lab or manufacturing hardware, and the same gap exists elsewhere ^[7]. Anthropic's own rollout is deliberately cautious in response: MHS remains a limited, application-only research preview, and the company plans to run safety evaluations with partners before eventually open-sourcing the standard with operational guidelines once the preview concludes ^[7]. That caution sits awkwardly next to the pace of the technical results - a lab that can go from driver-writing to a finished experiment in eight hours, or a laser controller that holds lock for 19 hours unattended, is moving faster than the rules meant to govern it, which is precisely the tension regulators will have to resolve before agents like this move from research preview to production hardware.