Previewing the Model Hardware Standard

TL;DR

The Model Hardware Standard (MHS) is Anthropic’s proposed shared specification for AI agents to safely discover, describe, and operate physical devices — lab instruments, robot arms, cameras, liquid handlers, plate readers — through a standardized driver exposing simple read/write primitives plus a natural-language tag layer that carries device characteristics (weight, safety limits, sensor semantics) the code alone doesn’t reveal. It is model-agnostic, agent-harness-agnostic, and accessed via standard protocols including MCP; the launch drops it into a research preview with partners (Genentech, HHMI Janelia, QuEra, CMU, UW Baker/Pinglay, Tetsuwan) plus vendor integrations (AWS Strands Robots, Automata, Danaher, Doosan, MBF Bioscience, QIAGEN, Tecan, Universal Robots, Hugging Face LeRobot, Raspberry Pi) rather than as an open-source release. The design bet is that hardware integration — currently weeks-to-months of bespoke work per lab — collapses to hours or minutes once devices become self-describing and self-discoverable to agents.

Key claims

MHS introduces a standardized driver that translates between an OS and a hardware device using a small primitive set (read, write) plus discovery in a common format, so devices and agents find each other across networks without bespoke translator programs [§How MHS works].
The driver carries natural-language tags — device weight, safety limits, “what can be measured / adjusted” — that automatically produce a reference file the agent uses to operate a device it has never seen before, replacing paper manuals and tacit knowledge [§How MHS works].
Agents control hardware through three mechanisms in parallel: MCP, the command line interface, and code files (APIs), enabling orchestration across multiple devices via a single line of code [§How MHS works].
Under the standard, an agent chains driver commands from one or more devices in code files to run long-running or fast operations without reasoning at every step, while still supervising and adjusting parameters as conditions change [§How MHS works].
Reported quantitative outcomes from partner deployments: QuEra reports a Claude-driven controller that recovers laser lock 99.3% of the time without human intervention on their neutral-atom quantum computers [§Early examples]; Carnegie Mellon reports ~3× faster serial-dilution dose-response experiments coordinating a liquid handler + plate reader + robot arm + monitoring cameras across three computers with incompatible interfaces [§Early examples]; HHMI Janelia reports MHS unifying a microscopy rig that previously required seven different vendor programs with no shared interface [§Early examples].
Explicit failure mode disclosed: as an LLM trained on text and images, Claude has limited physical intuition — Genentech researchers had to guide Claude to recognize foaming-in-samples as a physical failure rather than a software bug, so expert oversight remains required [§Joining the research preview].
MHS is not yet open-sourced; the research preview is scoped to safety-evaluation building, best-practice development, and expansion to hardware currently lacking a programming interface, with an eventual open release planned [§Joining the research preview].
MHS’s origin is a collaboration between Anthropic’s Beneficial Deployments team and Arco Bast (HHMI Janelia), who built a shared-memory dictionary allowing lasers, motorized focusers, and specialized cameras to communicate at memory speed on a brain-imaging rig — MHS generalizes that private interface into a public spec [§Acknowledgments].

Method

MHS is a specification, not a model. Its concrete artifact is a driver interface with four cooperating pieces. First, a primitive layer — every device exposes itself through read (get temperature, get frame, get joint state) and write (set temperature, move to, dispense volume) commands regardless of vendor. Second, a discovery format so devices and agents advertise their presence and capabilities on a network in a common structure. Third, a tag layer where users (or an agent interviewing them) write natural-language annotations describing device characteristics, safety limits, and semantics that the code interface doesn’t reveal; the driver compiles these into a reference file the agent consumes before operating the device. Fourth, an agent control surface exposing the device to the agent through three modalities: MCP for interactive tool calls, a command-line interface for scripting, and code files (APIs) for chaining commands into deterministic sequences that run faster than online reasoning would allow.

The reported operational pattern is that Claude explores a device empirically — makes an adjustment, observes results through the device’s own sensors, iterates — then codifies what it learned as a deterministic script it runs as a single command later. The laser-alignment example is treated as canonical: Claude interactively probed how each adjustment moved the beam through the camera, then packaged the aligned procedure as code so the whole process could subsequently run without step-by-step reasoning. This “explore then compile” loop is what the shared-memory-dictionary origin at HHMI Janelia optimized for.

Results

Anthropic reports six partner deployments as MHS’s evidence base at preview launch:

Genentech: MHS used to automate the BCA protein assay across a liquid handler, robotic arm, and plate reader; treated as a proof of concept for lab automation.
UW Baker/Pinglay labs: Zihao Song built a remote monitoring dashboard, an agent-supervised qPCR that halts on amplification curves, and a collision-free robot-arm / liquid-handler handoff.
Carnegie Mellon: ~3× faster serial-dilution dose-response experiments orchestrating four device classes across three computers with incompatible native interfaces.
HHMI Janelia (Ahrens lab): One microscopy rig that previously ran under seven vendor programs unified under a single MHS-orchestrated interface.
QuEra Computing: Agent-authored controller recovers neutral-atom-computer laser lock 99.3% of the time autonomously.
Tetsuwan Scientific: MHS integrated with the ResearchOS platform to run a citizen-science qPCR workflow characterizing San Pedro Creek pollution.

Vendor / library integrations announced: AWS Strands Robots (private pre-release for the preview), Automata LINQ, Danaher smart instruments, Doosan robotic arms, MBF Bioscience ScanImage, QIAGEN QIAsymphony Connect, Tecan Fluent, Universal Robots, Hugging Face LeRobot, Raspberry Pi (including a Camera MHS Driver). No standardized benchmark numbers, adoption counts, or ablations are reported — the post positions MHS as a partnered research preview rather than a benchmarked release.

Why it’s interesting

MHS is the first filed hardware-facing tool-use protocol from a frontier lab on the wiki, extending the ReAct-over-MCP substrate catalogued in Tool-Use Agents from digital tools (browsers, code interpreters, MCP servers over APIs) into the physical world — the “tools” here are microscopes, robot arms, and plate readers, and the agent needs to reason about safety limits and device characteristics that pure code interfaces don’t expose. It complements DimOS — The Agentive Operating System for Physical Space (v0.0.13) (DimOS’s @skill-decorated robot Modules auto-exposed as MCP tools) and ROSClaw — Give AI Agents a Body That Learns (ROSClaw’s Body/Sandbox/Practice runtime with mandatory sandbox validation before hardware contact): DimOS and ROSClaw are third-party open runtimes trying to become the standard from the framework side; MHS is a frontier lab writing the standard from the model side with vendor buy-in already lined up. The “explore then compile to deterministic script” operational pattern is the physical-world analog of Training-Free Group Relative Policy Optimization‘s learned-experience-library recipe on the Tool-Use Agents page — Claude reduces required interaction depth by encoding empirically-discovered device semantics into reusable code, rather than by increasing the tool-call budget in the style of MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling. It also sharpens How Claude Performs on Robotics Tasks‘s “Claude as VLA supervisor” thread — that paper treated VLA-outputted 7-DoF actions as the tool; MHS reframes the tool as the device itself one abstraction layer down, so the agent can operate hardware that has no VLA behind it at all. Contrasts with Addressing the Orchestration Gap in Generalist Robots via Physical Agency (Pigey)‘s Pigey (VLM orchestrator over TAMP + frozen VLA in a lab-benchmark setting) by targeting deployment across scientific-research and manufacturing settings where a shipped standard matters more than a benchmark score. The concrete open-question surface: MHS’s tag layer is a natural-language contract between a user and an agent about what a device is — how brittle is that contract under adversarial tag content, and does the “no physical intuition” limitation (Genentech’s foaming example) generalize to more subtle failure modes vendors haven’t seen? MHS ships with vendor commitments but no safety-evaluation numbers yet; the research preview is explicitly framed as the phase where those get built.