The Model Hardware Standard Directory
Keep up to date on protocol updates, new hardware integrations, and who is shipping.
This is an independent community project. It is not affiliated with, endorsed by, or supported by Anthropic in any way.
How it fits together
AI models power agents that discover and safely control physical devices through the Model Hardware Standard. MCP is one software interface agents may use to call an MHS driver — complementary tools, not competing specs.
What this is
The Model Hardware Standard (MHS) is Anthropic’s research-preview specification, announced 27 August 2026 with HHMI Janelia, for AI agents to discover and safely operate programmable physical devices. A standard driver uses read/write primitives, network discovery, and natural-language safety tags. Agents reach devices over MCP, a CLI, or generated code files. Access is apply-only; open source is planned after safety work.
MHS Directory is an independent community index of Model Hardware Standard devices, vendors, labs, and drivers. It is not affiliated with or supported by Anthropic. Anthropic’s spec site is apply-only; this catalog lists named public partners (Genentech, Tecan, Universal Robots, Hugging Face LeRobot, Raspberry Pi, AWS Strands Robots, Waters, and others).
What is MHS · FAQ · Anthropic announcement · 2026-10-29 news · Momentum
Subscribe for updates
Spec drops, new drivers, catalog changes. No ads. We email when something actually shipped.
Email me when the spec is public.
Apply-only today. We'll ping once if/when it opens. Not ship-news.
Featured
Official partners and key projects
AWS will support MHS through Strands Robots, a public library for connecting agents to physical devices. The 1.0 roadmap says every Robot is an MHS device and describes mounting robots on an MHS broker; that support is not in the public repo yet.
Adding MHS support in LeRobot so open robotics stacks can speak the same driver/discovery model. Watch this repo for the first public driver-shaped code.
Hot
Moving now
Adding MHS support in LeRobot so open robotics stacks can speak the same driver/discovery model. Watch this repo for the first public driver-shaped code.
AWS will support MHS through Strands Robots, a public library for connecting agents to physical devices. The 1.0 roadmap says every Robot is an MHS device and describes mounting robots on an MHS broker; that support is not in the public repo yet.
The open-source alternative to Anthropic's Model Hardware Standard (MHS). A vendor-neutral driver specification for AI-controlled hardware. Apache-2.0 licensed.
Open Claude Agent Skills for ICS/OT engineers — ladder logic, Modbus, OPC UA, and incident triage. Created for Anthropic's Model Hardware Standard (MHS). MIT-licensed.
Open Model Hardware Standard - AI agent hardware control. MIT-licensed community implementation from South China University of Technology ESA.
Catalog
18 public listings. Community-compiled from Anthropic’s announcement and follow-on coverage.
Status legend
18 of 18
- Hugging Face LeRobotHugging Facefeaturedbuildingplatform
Adding MHS support in LeRobot so open robotics stacks can speak the same driver/discovery model. Watch this repo for the first public driver-shaped code.
- AWS Strands RobotsAmazon Web Servicesfeaturedbuildingplatform
AWS will support MHS through Strands Robots, a public library for connecting agents to physical devices. The 1.0 roadmap says every Robot is an MHS device and describes mounting robots on an MHS broker; that support is not in the public repo yet.
- Automata LINQAutomatabuildingvendor
Adding MHS support to LINQ for intelligent error handling of instruments in autonomous labs.
- Carnegie Mellon — dose-response automationCarnegie Mellon Universitypreview userlab
Serial dilution dose-response ~3× faster. Agent orchestrated liquid handler, plate reader, robotic arm, and monitoring cameras across three computers with incompatible interfaces.
- DanaherDanaherexploringvendor
Exploring MHS-supported capabilities for smart instruments and autonomous laboratories to scale biomedical R&D.
- Doosan RoboticsDoosan Roboticstestingvendor
Testing MHS with robotic arms, including automated QA and coordinating tasks across multiple robots.
- Genentech — BCA protein assay PoCGenentechpreview userlab
Proof-of-concept automating the BCA protein assay across a liquid handler, robotic arm, and plate reader. Researchers still had to teach Claude that foaming was a physical failure, not a software bug.
- HHMI Janelia Research CampusHoward Hughes Medical Instituteorigin partnerlab
MHS started here: Arco Bast’s shared-memory dictionary for a multi-vendor brain-imaging rig (lasers, motorized focusers, cameras). Virginie Ruetten (Ahrens lab) unified a rig that previously needed seven vendor programs.
- MBF Bioscience ScanImageMBF Biosciencebuildingvendor
Building an MHS driver for ScanImage (laser-scanning microscopes in hundreds of neuroscience labs) for agent-driven real-time analysis and experiments.
- Model Hardware StandardAnthropicresearch previewstandard
Shared driver spec: read/write primitives, network discovery, natural-language tags that generate a device reference file with safety limits. Model-agnostic; agents reach it via MCP, CLI, or code files.
- QIAGEN QIAsymphony ConnectQIAGENpocvendor
Working PoC: agents troubleshooting instrument issues, guiding operators through recovery, improving uptime while reducing sample risk.
- QuEra Computing — laser lockQuEra Computingpreview userlab
Agent control of parts of the laser system in neutral-atom quantum machines. Recovered laser lock 99.3% of the time without human intervention.
- Raspberry Pi Camera MHS DriverRaspberry Pitesteddevice
Successful tests of a Camera MHS Driver; Raspberry Pi enabling MHS integration across a number of products. Highest-leverage hobbyist/developer device to index first.
- Tecan FluentTecanbuildingvendor
Adding MHS support so agents can discover and operate Fluent liquid handling platforms directly.
- Tetsuwan Scientific ResearchOSTetsuwan Scientificpreview userplatform
MHS integrated with ResearchOS. qPCR workflow for citizen-science pollution profiling in San Pedro Creek, California.
- Universal RobotsUniversal Robotsearly accessvendor
Early access to MHS; plans to add support to its robotics platform.
- University of Washington Baker and Pinglay labsUniversity of Washingtonpreview userlab
Zihao Song: remote instrument dashboard; agent-supervised qPCR that watches amplification curves and halts on time; collision-free plate handoff between arm and liquid handler.
- Waters CorporationWaters Corporationbuildingvendor
Joined Anthropic's MHS research program so agents can operate Waters LC-MS instruments. Also launched Waters AI Link, which connects waters_connect to Claude via MCP (early access now; general availability planned for November).
Community / Unofficial
Independent inspired-by work. Not Anthropic partners. Not the official spec.
- A3S MHS LaboratoryunofficialA3S-Lab
Deterministic virtual hardware laboratory for testing A3S agents at the MHS boundary. Conveyor workcell simulation with safety interlocks, emergency-stop, fault injection, and deterministic state. Explicitly a simulated MHS surface, not a conformance claim.
- Agnes MHS (Agnes Harness)unofficialAgnesAI-Labs
Independent same-name device protocol — not Anthropic's MHS, not affiliated with or endorsed by Anthropic, and makes no compatibility claim. Agnes Harness (Apache-2.0, ~1.28K stars) ships its own Agnes MHS 1.0 plus MOS 1.0 (Model Observation Standard): devices (robots, vehicles, sensors) dial AgnesHub, register tools/data sources, take calls and stream data; the model acts only through the hub. JSON Schema mhs/v1, mhs-check conformance profiles (Core, Motion, Manual, Pause, Perception, Maps, Streaming, Video, Audio clip), Python and TypeScript device libraries, sample devices/dev-hub (no hardware needed), and an optional Agnes plugin with seven device tools (list_devices, read_device, call_device, set_device, stop_device, watch_device, unwatch_device). README: "Agnes MHS is independent of Anthropic's standard of the same name."
- edge-mhs (Fastly)unofficialFastly
Fastly Compute edge security gateway for MHS device control over MCP — JWT auth, schema validation, hardware safety-limit checks, per-device quotas, audit logging (forked from edge-mcp).
- EVA-ClientunofficialNoietch
EVA-Client: A Unified Framework for Deployment, Evaluation, and Data Collection on Real Robots. Apache-2.0 community project (~303 stars, created 2026-07-02). EVA-MHS Preview (2026-08-28) on the dev/mcp_preview branch explores an MHS-style agent-to-hardware interface — MCP service for coding agents to inspect runtime state, call policy models, and directly control robots. Not formally released. Unofficial / not affiliated with or endorsed by Anthropic.
- HardMeshunofficialhuzjie
AI hardware control mesh & lab automation platform - MHS (Model Hardware Standard) implementation: device abstraction, multi-protocol drivers, closed-loop calibration, safety guardrails, LLM agent control.
- MHS Bambu Lab printer (zhenwei2972)unofficialzhenwei2972
Community MCP server for Bambu Lab A1/P1/X1 printers over LAN, shaped around Anthropic MHS read/write channels, discoverable device descriptors, and driver-enforced safety limits (not official Anthropic MHS; no public spec yet).
- MHS Demounofficialsidman76
Demo agent implementing the Model Hardware Standard (MHS) read/write pattern against simulated lab hardware, via MCP + Claude.
- MHS ICS SkillsunofficialAshigo Research
Open Claude Agent Skills for ICS/OT engineers — ladder logic, Modbus, OPC UA, and incident triage. Created for Anthropic's Model Hardware Standard (MHS). MIT-licensed.
- MHS Interop Bridgeunofficialbahtoo
Independent open-protocol drivers, 23 fault-injection physical-world evals, and cookbooks inspired by Anthropic's public MHS description. SiLA 2, file-drop, ASTM/HL7 analyzer, and bedside-monitor drivers plus MCP (mhs_* tools) and CLI (mhs-bridge); instruments are simulated. Not official Anthropic MHS and makes no compatibility claim.
- MHS-Pi-Gatewayunofficial1005John
Unofficial Raspberry Pi device gateway billed as an early/shadow implementation aligned to the Model Hardware Standard — not Anthropic's MHS, not affiliated with or endorsed by Anthropic, and not an implementation of the gated official spec. Peripherals are described by YAML Device Capability Descriptors; cloud agents control them over MCP (streamable HTTP); a local rule engine closes the loop offline; a seven-stage guard chain (whitelist, param schema check, e-stop, L2 confirmation, rate limit, dedupe/mutex, liveness probe) with L0/L1/L2 risk tiers; 12-field JSONL audit log. Includes ESP32-S3 LED reference firmware (MSP-1 serial protocol), a v0.9 shadow spec, test cases and test reports. Spec doc says Anthropic's MHS spec is not public yet, so this maps to MHS's public ideas and plans a mapping layer to migrate once the spec is published. Chinese-language README. MIT.
- mhsbenchunofficialbluitz
Driver-enforced safety boundary and human-checkpoint run loop for AI-operated lab instruments (simulated). Modeled on Anthropic MHS / Genentech-style pilot — inject faults, steer the agent, freeze learned protocols. Unofficial / not affiliated with Anthropic.
- Model Hardware Standard Notes (topsun-bot)unofficialtopsun-bot
Notes on Anthropic Model Hardware Standard (MHS): AI agents ↔ physical devices. Chinese + English notes tracking Anthropic announcement — explicitly not the official spec. Community research/documentation repo.
- Open MHS (SCUT-ESA)unofficialSCUT-ESA
Open Model Hardware Standard - AI agent hardware control. MIT-licensed community implementation from South China University of Technology ESA.
- Open MHS (tongriyaotxt)unofficialtongriyaotxt
Open Model Hardware Standard - AI agent hardware control. MIT-licensed community implementation inspired by Anthropic's MHS.
- Open-MHS (Abenor Labs)unofficialAbenor Labs
The open-source alternative to Anthropic's Model Hardware Standard (MHS). A vendor-neutral driver specification for AI-controlled hardware. Apache-2.0 licensed.
- open-mhs (vraj130)unofficialvraj130
Open-source reconstruction of MHS concepts: manifest-described devices, read/write/invoke primitives, safety limits enforced below the agent. Includes simulated + hardware (Pico 2) thermostats, audit logging, three rings of defense. CLI + MCP + Python API.
- OpenHALunofficialsohan-bhat
Unofficial open-source protocol for AI agents to control hardware — not affiliated with Anthropic or MHS, not an implementation of the gated official spec. One .hal.yaml device file format with person-confirmed units and limits; a safety kernel checking range, rate, interlocks and confirmation before any driver acts; single-use grants per command; one stop puts every device in a safe state; hash-chained audit log. v0.1 in development with ~34 drivers (SCPI, Moonraker/Klipper, LeRobot arms, PyLabRobot, SiLA 2, OPC UA, ROS 2, ESPHome, etc.), all tested on simulator only. `openhal init` prints setup for Claude Desktop, Claude Code or Cursor (MCP). README says MHS is a limited research preview and it plans to stay compatible via a mapping doc (spec/MHS-MAPPING.md) and a future adapter once a spec is published. Python, Apache-2.0.
- OpenMHS (KNambiarDJsc)unofficialKNambiarDJsc
Independent, model-agnostic supervisory gateway between AI agents and programmable hardware; draft spec v0.1 with JSON Schemas, deterministic policy (grants, approvals, unit/range, rate/slew limits), command engine with idempotency, UNKNOWN_OUTCOME handling and readback verification, hash-chained audit log, MCP server and Python SDK; simulated devices only (no real-hardware adapter yet). Not official Anthropic MHS and makes no compatibility claim.
- OpenMHS (Pratyush)unofficialPratyush-exe
OpenMHS is a shared specification that lets any language-model discover & safely operate hardware. Community-driven implementation inspired by Anthropic's Model Hardware Standard.
- Quilt MHSunofficialSuperInstance
quilt × Anthropic's Model Hardware Standard: quilt cells drive MHS devices; quilt runtimes exposed AS MHS devices. Runs today on MockMHS + conformance suite.
- Qwen-MM-Plugins MHSunofficialQwenLM
Qwen multimodal agent plugins capability for real hardware via Model Hardware Standard adapters — mhs_discover / mhs_meta_info / mhs_read / mhs_write / mhs_health_check / mhs_reset as MCP tools, with host-side safety limits, emergency-stop, and a bundled mock adapter. Unofficial / not Anthropic MHS.
- TARE (Trusted Adapter for Readback-verified Equipment)unofficialMrBesterTester
Independent design inspired by Anthropic's public MHS description: AI agents operate a virtual lab (microscope, mixed-signal oscilloscope, robot arm, liquid handler) via discover, configure, measure, and operate, with validation, instrument readback, and audit checks outside the model. Not official Anthropic MHS and makes no compatibility claim.
- Vision–Tactile Assembly Assistantunofficialbell-312
Unofficial robot assembly assistant in MuJoCo simulation (Franka Panda) — not Anthropic's MHS, not affiliated with or endorsed by Anthropic. Combines RGB coarse positioning, six-axis force/wrench feedback and contact-aware skills (visual approach, surface probing, plane alignment, spiral search, compliant insertion, retreat/retry/abort) in an observe–decide–act–verify agent loop. Built on Agnes Harness (AgnesAI-Labs/agnes-harness); registers the simulated robot as an MHS device via AgnesHub (vision_device.py), with camera/wrench via Agnes's Model Observation Standard (MOS). The 'MHS' it uses is Agnes MHS, which the Agnes repo states is independent of Anthropic's standard of the same name (see unofficial catalog entry Agnes MHS). Simulation only. English + Simplified Chinese READMEs. Apache-2.0. GitHub user bell-312 (no name/org given).