--- title: Anthropic wants Claude to run your robots, and the demos... description: Anthropic is previewing the Model Hardware Standard (MHS), a shared interface layer letting AI agents like Claude control physical hardware such as lab... canonical: https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Anthropic wants Claude to run your robots, and the demos are working | daily.dev og:description: Anthropic is previewing the Model Hardware Standard (MHS), a shared interface layer letting AI agents like Claude control physical hardware such as lab... og:url: https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5 og:image: https://api.daily.dev/og/posts/UtT30N9E5.png og:image:alt: Anthropic wants Claude to run your robots, and the demos are working og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Anthropic wants Claude to run your robots, and the demos are working **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 0 comments ## Summary Anthropic is previewing the Model Hardware Standard (MHS), a shared interface layer letting AI agents like Claude control physical hardware such as lab equipment, factory machines, and robots, without weeks of custom integration. MHS provides common drivers exposing read/write commands, device capability descriptions, and enforced safety limits, with agents connecting via MCP, CLI tools, or code. Hugging Face's LeRobot team already demoed Claude Code controlling a real SO-ARM101 robot arm. The safety-limit implementation details remain undisclosed, raising questions about how those limits are audited and overridden. ## Content Anthropic just launched a research preview of the Model Hardware Standard (MHS), and the headline number is almost unfair: QuEra Computing used Claude through MHS to automate laser-lock recovery on a quantum computer, cutting a 150-second, 58%-success manual process down to six seconds at 99.3% success. The AI outperformed a human specialist on lock-quality tuning across 695 of 700 trials. The idea is straightforward. Most lab and factory hardware speaks its own proprietary language, so wiring up a multi-vendor setup can take weeks. MHS standardizes that with simple read/write drivers plus reference files describing device capabilities, physical properties, and hard safety limits. Agents reach the hardware through MCP, command-line tools, or code. Carnegie Mellon and University of Washington researchers reportedly got multi-vendor hardware talking to each other in hours instead of weeks. It's positioned explicitly as a hardware counterpart to Anthropic's Model Context Protocol, and it's model-agnostic. The vendor list already includes AWS, Danaher, Tecan, QIAGEN, Doosan Robotics, Universal Robots, Hugging Face, and Raspberry Pi. Anthropic plans to open-source it after the preview. The QuEra case is worth sitting with. The agent didn't stay in control at runtime — it experimented autonomously on a dedicated testbed and wrote conventional software that then ran the recovery process. That's a meaningful distinction. It's not an AI with its hands on the laser at all times; it's an AI that figured out the procedure and handed off a working program. The honest limitation: agents still struggle when something goes physically wrong. Bubbles in a sample, a hardware failure, anything that requires actually looking at the thing — the understanding is programmatic, not physical. That's a real ceiling. There's also a regulatory wrinkle worth flagging. Europe's Machinery Regulation 2023/1230 starts covering AI-based safety functions in machinery from January 2027. MHS files that constrain robot arm speeds and angles could end up classified as regulated safety components. Anthropic is building something that might need to pass a compliance audit before it touches a factory floor. Lydia Hallie is already asking who's using Claude Code for firmware, robots, and home automation. The community is clearly curious. Whether MHS becomes the USB-C of lab hardware or another abandoned standard depends entirely on whether the vendor list keeps growing after the preview ends. ## Questions this post answers ### What is Anthropic's Model Hardware Standard (MHS) and how does it work? MHS is a research preview from Anthropic that lets AI agents like Claude control lab and factory hardware through standardized read/write drivers plus reference files describing device capabilities, physical properties, and safety limits. Agents access hardware via MCP, command-line tools, or code. It's positioned as a hardware counterpart to Anthropic's Model Context Protocol and is model-agnostic, with vendors including AWS, Danaher, Tecan, QIAGEN, Doosan Robotics, Universal Robots, Hugging Face, and Raspberry Pi. _Engineers evaluating agentic hardware control can follow MHS developments as they unfold on daily.dev._ ### How much did Claude improve laser-lock recovery on QuEra's quantum computer using MHS? Claude cut QuEra Computing's laser-lock recovery process from 150 seconds at 58% success down to 6 seconds at 99.3% success, outperforming a human specialist on lock-quality tuning across 695 of 700 trials. Notably, the agent didn't control the hardware live at runtime; it experimented autonomously on a dedicated testbed and wrote conventional software that then executed the recovery process. _Anyone weighing AI agents for hardware automation can track real-world results like this via daily.dev._ ### What are the current limitations of AI agents controlling physical lab or factory hardware? Agents still struggle when something goes physically wrong, such as bubbles in a sample or a hardware failure, since their understanding is programmatic rather than physical, requiring someone to actually look at the equipment. Additionally, Europe's Machinery Regulation 2023/1230 will start covering AI-based safety functions in machinery from January 2027, potentially classifying MHS safety files as regulated components needing compliance audits. _Teams building agentic hardware systems can watch for regulatory and reliability updates through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 4 discussions and 144 comments across x (as of 2026-09-13). **TL;DR:** Replies are dominated by enthusiastic hobbyists and builders sharing what they've already done with Claude on firmware, robots, sensors, and home automation, with only scattered notes of caution about safety limits and reliability. **Sentiment:** 70% positive · 25% mixed · 5% skeptical **The case for** - Many describe cutting hardware/firmware project time from weeks to hours across robots, drones, PLCs, and sensor rigs. - Some highlight that physical feedback loops (readings, hardware behavior) keep the model honest since it can't fake results. - One person notes a thin, fast-adapting interface could let agents replan quickly across different lab/factory setups. **The pushback** - A couple of replies note the model still hallucinates pinouts or struggles with byte-level debugging on STM32/FPGA hardware. - One reply stresses that safety boundaries need to be locked down before a unified interface is trustworthy. - One person flags real annoyance with guardrails needing another AI to finish firmware ports. **By community** - x (positive): An overwhelming flood of builders sharing hardware/firmware/robotics projects they've built with Claude, with only isolated cautionary notes about safety limits and occasional model errors. **Open questions** - Will the interface be thin/fast enough for agents to adapt to new instruments or safety limits within minutes rather than requiring fresh integration work? - How rigorously are safety limits enforced before agents are trusted with actual machinery? **Highlights** > @rohanpaul_ai The agent also has to switch context fast — lab bench, cleanroom, factory cell — and rewrite the plan when the instrument, the material, or the safety limit changes. MHS only matters if that interface is thin enough for the model to adapt in minutes, not after another integration > — [truetrue2355 on x](https://x.com/truetrue2355/status/2093072418562818550) > @rohanpaul_ai 统一接口是好事，但安全边界得先焊死 > — [ajs6888 on x](https://x.com/ajs6888/status/2093072635039244579) > @lydiahallie firmware and sensors are exactly where the agent loop gets interesting because the feedback is physical. the model can't fake a reading that didn't happen > — [rusabuilds on x](https://x.com/rusabuilds/status/2093032511974899946) > @lydiahallie LLM hallucinates pinouts, breaks my delta bot. > — [stevencheng on x](https://x.com/stevencheng/status/2093030526764605831) > @lydiahallie Opus 5 and Fable 5 has issues with debugging on my side. They struggle to spot byte level instruction errors on STM32 and FPGA while Sol on Codex solved it. I believe a more aggressive strategy on web search would help, but connecting the dots is where they struggles. > — [junzhi2002 on x](https://x.com/junzhi2002/status/2093048868833513641) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2093064270028431640) · 0 points · 8 comments - [x](https://x.com/Thom_Wolf/status/2093075291744665928) · 0 points · 0 comments - [x](https://x.com/yoheinakajima/status/2093087564471418950) · 0 points · 0 comments - [x](https://x.com/lydiahallie/status/2092996152023232782) · 1 points · 136 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#mcp](https://daily.dev/tags/mcp), [#robotics](https://daily.dev/tags/robotics), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic) [View this post on daily.dev](https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Anthropic wants Claude to run your robots, and the demos are working","url":"https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5"},"datePublished":"2026-08-27T20:37:13.099Z","dateModified":"2026-09-13T20:03:51.975Z","description":"Anthropic is previewing the Model Hardware Standard (MHS), a shared interface layer letting AI agents like Claude control physical hardware such as lab...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,mcp,robotics,claude,anthropic","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Anthropic wants Claude to run your robots, and the demos are working"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/anthropic-wants-claude-to-run-your-robots-and-the-demos-are-working-utt30n9e5#faq","mainEntity":[{"@type":"Question","name":"What is Anthropic's Model Hardware Standard (MHS) and how does it work?","acceptedAnswer":{"@type":"Answer","text":"MHS is a research preview from Anthropic that lets AI agents like Claude control lab and factory hardware through standardized read/write drivers plus reference files describing device capabilities, physical properties, and safety limits. Agents access hardware via MCP, command-line tools, or code. It's positioned as a hardware counterpart to Anthropic's Model Context Protocol and is model-agnostic, with vendors including AWS, Danaher, Tecan, QIAGEN, Doosan Robotics, Universal Robots, Hugging Face, and Raspberry Pi. Engineers evaluating agentic hardware control can follow MHS developments as they unfold on daily.dev."}},{"@type":"Question","name":"How much did Claude improve laser-lock recovery on QuEra's quantum computer using MHS?","acceptedAnswer":{"@type":"Answer","text":"Claude cut QuEra Computing's laser-lock recovery process from 150 seconds at 58% success down to 6 seconds at 99.3% success, outperforming a human specialist on lock-quality tuning across 695 of 700 trials. Notably, the agent didn't control the hardware live at runtime; it experimented autonomously on a dedicated testbed and wrote conventional software that then executed the recovery process. Anyone weighing AI agents for hardware automation can track real-world results like this via daily.dev."}},{"@type":"Question","name":"What are the current limitations of AI agents controlling physical lab or factory hardware?","acceptedAnswer":{"@type":"Answer","text":"Agents still struggle when something goes physically wrong, such as bubbles in a sample or a hardware failure, since their understanding is programmatic rather than physical, requiring someone to actually look at the equipment. Additionally, Europe's Machinery Regulation 2023/1230 will start covering AI-based safety functions in machinery from January 2027, potentially classifying MHS safety files as regulated components needing compliance audits. Teams building agentic hardware systems can watch for regulatory and reliability updates through daily.dev."}}]} ```