Background · not affiliated with Anthropic
Relation to Anthropic's MHS preview
Anthropic opened a research preview of its Model Hardware Standard (MHS) in August 2026. OpenMHS is a separate, independent project. This page summarises what the preview's public sources describe — the announcement, the landing page, partner write-ups and the two Anthropic videos — separates verified facts from claims that were never published, and states how OpenMHS relates to it.
How it is described
A standardized driver, tags, and three control surfaces
The public description is consistent across the announcement and the videos; the wire-level detail is not published.
1 · Standardized driver
“Software that translates between a computer’s operating system and a hardware device”, using simple primitives such as read (“get temperature”) and write (“set temperature”).
2 · Standard discovery
Devices become “discoverable in a standard format, so that devices and agents can find each other and communicate across networks” without a bespoke translator.
3 · Tags become a reference file
Users write natural-language tags (or an agent interviews them). The driver “automatically produces a reference file” describing what the device measures, what can be adjusted and which safety limits are enforced.
4 · Three mechanisms
“MCP, the command line interface, and code files (APIs). These work together to enable orchestration across multiple devices via a single line of code.”
5 · Code files for long work
The agent can “chain together driver commands from one or more devices in code files”, so hardware runs without the model reasoning at every step. Learning is captured as a deterministic script.
6 · Origin: shared memory
The collaboration began at HHMI Janelia with a shared memory dictionary that let instruments “communicate … at memory speed”. How this relates to network-wide discovery is not stated.
What the videos add
A short film and an 11-minute documentary
Video facts are cited as video facts. The short film is <code>UxJZrCFzTHY</code> (2:13); the documentary with the demonstrations is <code>P1zBiAQU1IA</code> (11:10), chapters 0:00 · 0:47 · 2:38 · 4:30 · 6:14 · 8:11. Timed captions were not retrievable, so timestamps are chapter-level or approximate.
| Demonstration | What it shows | Mechanism it illustrates |
|---|---|---|
| Janelia microscope | Custom multiphoton rig with lasers, focusers and cameras from different vendors; the first MHS layer is used to control stage and laser remotely. | Driver + read/write primitives on a heterogeneous rig. |
| Robotic arm safe range | Operator defines the arm’s safe range; an out-of-range command is refused by MHS, then a first-time grasp task is attempted. | Safety limits enforced below the agent (asserted; location not specified). |
| Leica microscope | Claude focuses, moves around the sample, changes magnification mindful of crashing into the sample, captures and interprets stained tissue. | Closed-loop read → decide → write on a precision instrument. |
| Diatom tracking | Claude writes a tracking script; after the first attempt fails it adds a live UI and tracks a swimming organism for minutes. | Agent-synthesized code file becoming the runtime controller. |
| Genentech liquid handling | Claude runs transfers, reads absorbance, detects bubbles in wells and adjusts flow-rate parameters in a closed loop. | Multi-device read → decide → write; error recovery with human guidance. |
The videos also use marketing language (“recovered even overnight”, “first time in history”) that the written material qualifies: Genentech reports that bubble failures needed human guidance because the model first treated a physical failure as a software retry.
Safety: asserted vs specified
A strong claim, a thin specification
The preview asserts enforcement below the agent; it does not publish the enforcement model.
| Claim | Source | What is stated |
|---|---|---|
| Standardized driver with read/write primitives | Announcement | A driver translates the OS to a device and exposes read/write-style primitives. Wire format not stated. |
| Standard discovery across networks | Announcement | Devices and agents find each other in a standard format. Protocol and scope not stated. |
| Tags → automatically generated reference file | Announcement | Tags in natural language produce a file describing measures, adjustments and enforced limits. Format not stated. |
| Three control mechanisms | Announcement | MCP, CLI and code files work together for orchestration via a single line of code. |
| Code files chain driver commands | Announcement | Long-running work runs in code files without per-step model reasoning. |
| Shared-memory dictionary at memory speed | Announcement (origin) | The Janelia origin was a shared memory dictionary; its relationship to network discovery is not stated. |
| Safety enforced at the hardware interface | QuEra write-up | Device-declared bounds, interlocks and e-stops enforced independently of the model. Enforcement code location not specified. |
| Out-of-range arm motion refused | Documentary (chapter 4:30) | After an operator defines a safe range, an out-of-range command is shown being refused. Timestamp cannot be re-derived (timed captions blocked). |
| Six induced fault conditions blocked | CMU case study | Missing plate, e-stop and other conditions blocked before motion. Six conditions are described in prose. |
| Devices without a programmable interface unsupported | Announcement | Explicit current limitation; GUI-only devices are driven through GUI automation instead. |
Early ecosystem
Who was named, and how strongly
Modality matters: case studies are proof-of-concept, pilots and demos; vendors range from exploring and testing to adding and building.
| Partner | As described in the announcement |
|---|---|
| Genentech | proof-of-concept BCA assay across liquid handler, robotic arm and plate reader; closed-loop flow-rate optimisation |
| HHMI Janelia | origin rig, shared-memory dictionary, live microscope control |
| QuEra | laser relock and PID tuning on quantum hardware; strongest written safety statement |
| Tetsuwan | ResearchOS replaced its scheduler and driver stack with MHS commands; one connector per instrument |
| CMU | six instruments, drivers built with an agent exploring vendor interfaces, GUI-only plate reader |
| Danaher / Leica | actively exploring MHS support for smart instruments |
| QIAGEN | working proof-of-concept on QIAsymphony Connect |
| Tecan | adding MHS support for Fluent liquid handlers |
| Universal Robots | early access; plans to add support |
| MBF Bioscience | building an MHS driver for ScanImage |
| AWS | will support MHS through Strands Robots (private pre-release) |
| Automata | adding MHS support to LINQ |
| Doosan Robotics | testing MHS with robotic arms |
| Hugging Face | adding MHS support in LeRobot |
| Raspberry Pi | enabling MHS integration after Camera MHS Driver tests |
How OpenMHS relates
Adopt the framing, define the contract, reject the unsafe shortcuts
OpenMHS 0.9 starts from the same problem statement and is independent of the preview: it is not a reimplementation, a compatible dialect, or an endorsed continuation.
Adopted
- Integration-cost framing — Bespoke integration is the bottleneck; a shared driver layer removes the N×M translator problem.
- Physical knowledge as data — A device's physical characteristics belong in machine-readable metadata, not in manuals or tacit knowledge.
- Natural-language tags — User-authored tags are first-class data; an agent may interview the operator to elicit them.
- Generated reference view — Every device gets a generated, model-readable view of what it measures, adjusts and enforces (here: the device profile).
- Three control surfaces — MCP, a CLI and a code API are used together (here: transport bindings T4/T5/T6).
- Code files for long work — Chaining commands into deterministic code escapes the per-step reasoning loop (here: versioned compositions with child receipts).
Defined here
- Wire protocol — JSON-RPC 2.0, framing, deadlines, cancellation and a stable error taxonomy.
- Receipts and evidence — Command ids, revisions, canonical state hashes, hash-chained evidence verifiable offline.
- Enforced safety envelope — Risk classes, limits, interlocks, single-use permits bound to the input digest, latching e-stop.
- Simulation as a mode — A mandatory declared fidelity contract; no silent sim/hardware switching.
- Adapter and bridge model — Declarative manifests plus reusable bridges for existing SDKs and simulators.
- Executable conformance — CL0–CL5, suite-produced digest-addressed reports instead of prose claims.
Deliberately rejected
- Free-form writable global state dictionary — breaks least privilege and auditability; mutations stay capability- and receipt-scoped.
- Shared memory as the core transport — OpenMHS is a supervisory contract, not a real-time bus; memory-speed is a deployment property.
- Auto-published model-authored drivers — registry admission requires review, evidence and trust.
- Model-side safety claims — OpenMHS conformance is endpoint behaviour; model evaluations are outside the specification.
Claims this project will not make
Boundaries that keep the analysis honest
The preview is closed, apply-only and has no public wire specification; no artefact here may imply otherwise.
- Compatibility with, or endorsement by, Anthropic or the MHS research preview.
- Implementing the preview's (unpublished) wire format or driver API.
- Having received preview access, or speaking for any named partner.
- That OpenMHS conformance is a safety evaluation or a safety certification.
- That any vendor named in the preview supports OpenMHS.
- That “memory-speed” or “bare-metal speed” is an OpenMHS guarantee.
Sources
Fetched and checked on 2026-10-20
Per-claim quotes, transcript status and gap analysis are kept with the specification sources.
- Anthropic — Previewing the Model Hardware Standard (Aug 27, 2026)
- The research-preview landing page operated by Anthropic (its address is cited in the appendix below; it is not this project's site)
- Short film — Model Hardware Standard: AI operating physical equipment (2:13)
- Documentary — AI models can now help run physical science experiments (11:10)
- Tetsuwan — integrating MHS into ResearchOS
- QuEra — holding the light: AI, lasers and MHS
- Appendix D — the relationship appendix, with per-claim sources