# Model Hardware Standard

> Source: https://aiwiki.ai/wiki/model_hardware_standard
> Updated: 2026-08-28
> Fact-checked: 2026-08-28
> Categories: AI Agents, AI Tools & Products, AI for Science, Anthropic
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Model Hardware Standard." aiwiki.ai, 28 Aug 2026. https://aiwiki.ai/wiki/model_hardware_standard
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Model Hardware Standard** (**MHS**) is an early specification for connecting AI agents to programmable physical equipment. [Anthropic](https://aiwiki.ai/wiki/anthropic) and the Howard Hughes Medical Institute's Janelia Research Campus began developing it for scientific instruments and advanced-manufacturing systems that otherwise require device-specific integrations. Anthropic opened MHS to a limited, application-based research preview on August 27, 2026. The specification was not open source or generally available at launch.[1][2]

MHS defines a common driver layer through which software can discover, read from, and write to equipment such as microscopes, liquid handlers, robotic arms, plate readers, and laser-control systems. It is intended to work with different models and agent harnesses, including connections made through the [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol) (MCP), a command-line interface, or programmatic APIs.[1]

## Status and scope

| Field | Detail |
| --- | --- |
| Public announcement | August 27, 2026[1] |
| Release state | Limited research preview, with access by application[1][2] |
| Origin | Collaboration between Anthropic and HHMI Janelia Research Campus[1] |
| Intended domains | Scientific research and advanced manufacturing[1] |
| Supported equipment | Devices with a programmable interface[1] |
| Model compatibility | Described as model-agnostic by Anthropic[1] |
| Public code at launch | Not released; open-sourcing was a future plan[1][2] |

The word "standard" describes the project's intended common interface. At preview launch, Anthropic had not identified an independent standards body, published a completed normative specification, or announced a finalized governance process. MHS should therefore be understood as a developing specification under a partner preview, not as an adopted industry standard.[1][2][3]

## Architecture

An MHS driver translates between a device's existing programming interface and a common representation. Anthropic describes two basic operation types: reads, such as retrieving a temperature, and writes, such as setting a temperature. Devices advertise themselves in a standard format so that software can discover them over a network.[1]

The driver can include natural-language tags that document physical characteristics, measurable values, adjustable parameters, and enforced safety limits. From this information, the system produces a reference file for an agent. Anthropic says an operator can write the tags directly or have an agent interview them about the equipment.[1]

MHS exposes equipment through MCP, command-line tools, and code interfaces. An agent can sequence operations, monitor measurements, and change parameters. For long-running or latency-sensitive work, it can also compose driver commands into conventional code rather than making every control decision through online model inference. MHS is therefore an equipment interface, not an AI model, an agent runtime, or a replacement for a device's underlying control software.[1]

## Safety model

The announced design places operating bounds in the device description and driver layer. In QuEra's pilot, engineers defined device bounds, interlocks, and emergency stops, and reviewed the agent's experimental scope and success criteria. The final laser-recovery controller was a deterministic, inspectable program rather than a model making production-time decisions.[1][4]

These mechanisms do not establish that arbitrary MHS deployments are safe. Anthropic said the preview would be used to develop additional evaluations, deployment practices, and a physical-safety roadmap. It also acknowledged that current models have limited spatial and physical reasoning, can mistake physical failures for software problems, and continue to require expert oversight. Equipment without a programming interface was unsupported at launch.[1]

Research on autonomous laboratories independently supports the need to separate model reasoning from constrained execution. Proposed systems have separated probabilistic model reasoning from deterministic execution through structured validation and command verification, with operational boundaries and explicit control mechanisms for safety-critical instrument control.[5][6] These papers did not evaluate MHS and do not validate its implementation.

## Preview partners

The initial ecosystem included research groups, equipment makers, robotics developers, and cloud or software providers. Anthropic named Amazon Web Services, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, and Universal Robots as organizations supporting, testing, exploring, or planning MHS integrations. It also said Hugging Face was adding support to LeRobot and Raspberry Pi was enabling integrations after tests of a camera driver.[1] The verbs differ materially: a planned integration, an exploratory project, and deployed support are not equivalent.

The launch announcement also presented demonstrations from Genentech, the University of Washington Baker and Pinglay labs, Carnegie Mellon University, HHMI Janelia, QuEra Computing, and Tetsuwan Scientific. Most results were reported by Anthropic or the participating organization and were not peer-reviewed independent evaluations.[1][4]

## Demonstrations

### LeRobot robot-arm demonstration

LeRobot, a [Hugging Face](https://aiwiki.ai/wiki/hugging_face) robotics project, reported a research-preview demonstration in which Claude Code controlled a physical SO-ARM101 arm through MHS. The team said the task used no learned policy, teleoperation, or recorded demonstrations. Instead, the agent measured the workspace and wrote motion commands. Calibration still required a person: software displayed 16 points in the camera view, the arm's torque was reduced, and a person placed the closed gripper on each point while the arm recorded the corresponding position.[1][7]

The team reported 4.1 mm position accuracy and 3.0 mm placement in that setup. Its best run placed 12 bricks while forming four color groups. These were single-team measurements from one demonstration; 12 bricks described the best run, not a success rate or repeated-trial result.[7] LeRobot said a fully hands-off run remained future work, while Anthropic described public LeRobot support as being added and the broader open-source MHS release as forthcoming.[1][7] LeRobot's standard SO-101 documentation treats calibration, teleoperation, data recording, and evaluation as separate workflows.[8] The case therefore does not establish learned-policy performance, autonomous calibration, generalization to other tasks or arms, or compliance with an external safety standard.[1][7][8]

### Laboratory automation

Genentech used a proof of concept spanning a liquid handler, robotic arm, and plate reader for a BCA protein assay. In the reported tests, [Claude](https://aiwiki.ai/wiki/claude) initially used a generic liquid-flow setting that caused bubbles in viscous samples. It later optimized separate flow rates against expert-prepared transfers, but researchers still had to explain that foaming was a physical problem requiring a physical correction. The demonstration showed both closed-loop parameter search and a limitation in the model's physical reasoning.[1]

University of Washington researchers connected six instruments in less than a week for demonstrations including qPCR supervision and robotic handoffs. They reported collision-free handoffs during repeated tests, while cautioning that complex protocols would require further optimization and that agent compute cost remained a concern.[1]

Carnegie Mellon researchers demonstrated a serial-dilution workflow. Their report said a system using Claude Opus 4.8 developed drivers and orchestration in about eight hours, blocked six deliberately induced safety conditions before robot motion, and revised a failed dose-response experiment. The work used a dye as a stand-in for a drug candidate and remained a proof of concept.[1]

Tetsuwan Scientific connected MHS to its ResearchOS lab platform for a qPCR workflow. It reported camera-based detection of liquid-handling errors, cross-device recovery using a centrifuge, and an optimization study covering 9,143 dispenses. Its held-out comparison concerned the company's own precision-prediction model and a manufacturer's technical specification, not a general benchmark of MHS.[1]

### Microscopy and quantum control

At HHMI Janelia, MHS grew from a shared-memory state dictionary used to expose multiple microscope components through one interface. Researchers described reusing analysis and visualization code across data streams, integrating a new camera in minutes, and developing supervised agentic microscopy workflows. These are accounts from developers and early users, not controlled comparisons across laboratories.[1]

QuEra used MHS on a dedicated test system for a neutral-atom [quantum computing](https://aiwiki.ai/wiki/quantum_computing) laser. The company reported that an agent-developed deterministic recovery script restored the target lock in 695 of 700 timed trials across seven induced fault types. It reported recovery times of under six seconds for most faults and about 10 to 14 seconds for the hardest cases, compared with five to ten minutes for an expert. QuEra also reported a separate overnight parameter-tuning exercise.[1][4]

The QuEra measurements concern one testbed and a bounded set of disturbances. QuEra said the agent needed extensive context, sometimes paused for human approval, and could not troubleshoot physical hardware failures outside its programmatic view. The company planned to extend the recovery approach to live processors and other subsystems, so those deployments were future work at announcement.[1][4]

## Evidence and limitations

MHS launched with unusually detailed partner narratives, but the public evidence had important limits. The central technical description and most numerical results came from Anthropic and participating organizations. No public conformance suite, interoperability certification, security assessment, complete threat model, or independent multi-site benchmark accompanied the preview.[1][3][4]

Some reported improvements compare MHS-assisted work with prior bespoke workflows, but the comparisons use different teams, tasks, equipment, and baselines. They do not demonstrate a universal reduction in integration time or performance improvement. Likewise, success on a laser-control testbed or a laboratory proof of concept does not establish safe operation of other equipment.[1][4]

The preview was designed partly to address these gaps. Anthropic said participants would help build safety evaluations and best practices before the project became open source. It promised to publish findings from the preview with deployment guidance when the code was released, but gave no date or license for that release.[1][2]

## References

1. Anthropic. "Previewing the Model Hardware Standard." Aug. 27, 2026. https://www.anthropic.com/news/model-hardware-standard-research-preview
2. Model Hardware Standard. "Model Hardware Standard." Research preview site, accessed Aug. 28, 2026. https://www.modelhardwarestandard.com/
3. Kyle Orland. "Anthropic's new hardware standard lets AI agents control the physical world." Ars Technica, Aug. 27, 2026. https://arstechnica.com/ai/2026/08/anthropics-new-hardware-standard-lets-ai-agents-control-the-physical-world/
4. QuEra Computing. "QuEra Computing Uses AI to Automate a Critical Quantum Computer Subsystem, Enabling the Acceleration of Commercial-Grade Quantum Computing Deployments from QuEra." Aug. 27, 2026. https://www.quera.com/press-releases/quera-computing-uses-ai-to-automate-a-critical-quantum-computer-subsystem-enabling-the-acceleration-of-commercial-grade-quantum-computing-deployments-from-quera
5. Zhuo Diao, Kouma Matsumoto, Linfeng Hou, Hayato Yamashita, and Masayuki Abe. "Autonomous Laboratory Agent via Customized Domain-Specific Language Model and Modular AI Interface." arXiv:2602.20669v1, Feb. 24, 2026. https://arxiv.org/abs/2602.20669v1
6. Zihan Zhang, Haohui Que, Junhan Chang, Xin Zhang, Hao Wei, and Tong Zhu. "Safe-SDL: Establishing Safety Boundaries and Control Mechanisms for AI-Driven Self-Driving Laboratories." arXiv:2602.15061, 2026. https://arxiv.org/abs/2602.15061
7. LeRobot. "We took part in the research preview of MHS from Anthropic." X, Aug. 27, 2026. https://x.com/LeRobotHF/status/2093074808141385788
8. Hugging Face. "SO-101." LeRobot documentation, accessed Aug. 28, 2026. https://huggingface.co/docs/lerobot/so101