Every machine is about to speak Claude

Friday's Rundown was about machines learning to speak Claude. I rewrote it the way I would tell a colleague over coffee. Short version: Anthropic wants a shared plug so AI helpers can run lab and factory kit without weeks of custom wiring. Developers in San Francisco are whispering to those helpers instead of typing. A named checklist in Claude Code interviews you when you are stuck. And a London team used live-video AI as a second pair of eyes in pituitary surgery — the piece a GP will actually care about.

The newsletter also carried ads for a LangChain Agentic Operating Model guide, a Gartner webinar, and a Switch sponsored tool. Skipping those.

Words worth knowing

Word	In one line
Agent	An AI helper that can take steps (look something up, run a machine), not only chat back.
MCP	Model Context Protocol — a shared plug so agents can talk to software tools the same way, instead of a one-off connector for each app.
MHS	Model Hardware Standard — the hardware version of MCP. A shared way for agents to find, learn, and run physical machines.
Research preview	Early access for a small group while they still work out the kinks. Not a finished product you can buy.
Open-source	The recipe is published so others can copy, inspect, and improve it. Anthropic says this comes later for MHS.
Custom code / API	One-off programs, or a private software doorway, written just for that machine. An API is the usual "please do this" doorway between programs.
Hugging Face	The big public library of downloadable AI models — and, here, a robotics toolkit (LeRobot) they are wiring into MHS.
Raspberry Pi	A cheap credit-card-sized computer people use for tinkering and small gadgets. They are adding MHS to device lines after camera tests.
Dictation	Speaking instead of typing. The computer turns speech into text.
Wispr Flow	A voice-to-text app some developers use to talk to their computer, including to agents.
iOS 27	Apple's next iPhone software. Apple has announced a rebuilt Siri AI and an AI dictation keyboard for it.
Terminal	The old-fashioned text window where you type commands instead of clicking.
Claude Code	Anthropic's coding assistant that lives in Terminal. You talk to it with typed commands, including slash commands like /unstuck.
Plugin / skill	A saved routine you install once and run later by typing something like /name. Think of a named checklist, not a new brain.

MCP started as a connector. It became the default way AI agents talk to software. Anthropic's new bet is the hardware version: the Model Hardware Standard (MHS), now in a research preview (early access, not a finished product).

The pitch: agents can look up, learn, and run factory and lab machines — microscopes, robotic arms, assembly lines — that previously needed custom code for each device. Anthropic wants MHS to spread as an industry standard the way MCP did for software.

Labs and factories usually spend weeks of specialist hand-wiring to get kit talking. Anthropic claims MHS can cut that to hours or minutes. Machine owners describe the equipment in ordinary language. MHS turns that into a reference file an agent can read: what the device measures, what you can change, what safety limits apply.

In one test, Claude taught itself to align a laser by trial and error — tweak, watch the beam on a camera, tweak again — then packaged what it learned into a script that did the job in a single pass, without reasoning at every step.

An open-source release is planned later, after the preview. Named partners include Tecan (lab liquid-handlers), QIAGEN (sample-prep instruments), and AWS (Amazon's cloud, via a robots library). Hugging Face is adding MHS to its robotics line (LeRobot). Raspberry Pi is adding it to device lines after tests with a camera driver.

Why a GP should even care

You will not plug MHS into the treatment room this year. The useful idea is the same as a shared pathology interface: once machines speak one language, an agent can run a workflow instead of a technician writing a one-off script for every microscope. That is a lab-automation story, not a clinic one. Treat the "hours not weeks" claim as Anthropic's.

Rowan's Corner: whispering to AI agents

Rundown writer Rowan Cheung's note: developers in San Francisco are using Wispr Flow (a dictation app) to control agents by voice. He thinks it is breaking out because Apple is shipping an AI dictation keyboard in iOS 27 plus a rebuilt Siri AI. Apple has announced both. The iPhone software itself is still rolling toward its usual autumn release, so this is a near-term product story, not something already on every phone.

His reasons, in short:

People speak at about 150 words a minute and type at about 60. Talking is faster.
Spoken prompts come out two to three times longer, so the agent gets more context.
A "director of agents" workflow: several agent tabs open, a 60-second whispered brief, review the result, next tab.
Dictation is now good enough to literally whisper in an office of 50 people.

His own habit: whisper ideas into Wispr Flow on walks. A scheduled Claude task organises them into Notion overnight.

How to start, per Rowan: wait for iOS 27, or map a Wispr-style app to the iPhone Action Button (the extra button on newer iPhones). Optional extra: a $20 desk microphone.

Practical: AI skills to make better decisions

The newsletter's how-to uses makerskills — a plugin (a pack of saved slash-command routines) for Claude Code. Honest bit first: most GPs will never install Claude Code. You do not need to. The idea still translates. It is a named checklist that interviews you when you are circling the same wall, or choosing between two options.

If you do live in Terminal, the install is two commands: open Claude Code, then /plugin marketplace add coreyhaines31/makerskills and /plugin install makerskills@makerskills.

Two skills worth knowing:

/unstuck — when you are circling the same problem. Example in the guide: a tool called social-fetch failed to pull an X (Twitter) post and suggested a paid scraping API (a paid doorway to grab the post). /unstuck asked whether the video even needed that demo.
/decide — when you are choosing between options. Example: it compared two hypothetical client projects, tested how reversible each choice was, recommended one, and set a date to revisit.

Pro tip

Run the skill inside a project that already has context — notes, files, the actual problem — so it is not interviewing a blank page. Same as bringing the chart to a case discussion instead of describing it from memory. If you never touch Claude Code, you can still steal the shape: a short list of questions you ask yourself before you buy a tool or pick a path.

UK live-video AI in brain surgery

This is the item a GP will actually remember. University College London Hospitals (UCLH) say they have done the first brain surgery with a live AI assistant. The system watched in real time through the surgical camera. It flagged hidden arteries and optic nerves to avoid while the tumour came out.

The patient was Rhys Hibbert. A seizure in late 2024 revealed an 11mm growth on the pituitary (the marble-sized hormone gland at the base of the brain). It slowly narrowed his field of vision. Surgeons reached the tumour through the nose — a standard endoscopic route for this gland, camera up the nose, not an open skull case.

The AI sat on its own monitor. It marked the safest spots to cut and the areas to leave alone. It was trained on hundreds of labelled videos of past pituitary removals. Surgeon Hani Marcus told the BBC it was trained on "more operations than most surgeons see or do in a lifetime" and can act like an expert second pair of eyes.

UCLH say the tumour came out and his vision was restored within days. The team is moving toward larger trials. The human surgeon stayed in control throughout.

News, not a practice-changing trial

This is one patient, in a specialist centre, in a research study. It should not change how you counsel someone with a pituitary incidentaloma tomorrow. The newsletter's own read is the right one: near-term this is a second set of eyes. Robots doing the actual movements is later. The surgeon still cuts. Do not add extra clinical claims on the back of a press release.

Community: a family heirloom catalogue

Reader Karen D. had a practical job: her mother moved into a retirement suite, and the family needed a catalogue of what was left. She numbered a grid on a whiteboard, put items in the squares, and took photos. Gemini (Google's assistant) identified the objects by grid square, categorised them, estimated market values, and built a spreadsheet with a signup column so relatives could claim pieces. Physical staging took 30–45 minutes. The digital pass took under 10 minutes.

Quick hits

OpenAI published an open letter signed by 116 companies, including Anthropic: AI-enabled cyberattacks will become far more widespread and sophisticated. It calls for coordinated cyber defence.
Nvidia is reportedly acquiring Hugging Face for $12.9 billion, per The Information. Rumour and report, not a closed deal — neither company has confirmed it.
Google Gemini Omni 1.1 Flash is an upgraded video model: 40-second scene extensions and 4K upscaling. It sits at the top of Arena's text-to-video board (the newsletter said text-to-image; the public Arena board is video).
Thinking Machines co-founder Barret Zoph is rejoining Google as a research vice president for reinforcement learning on Gemini, after a brief second stint at OpenAI.
Salesforce and Anthropic launched Claudeforce: a 37-skill sales plugin inside Claude, and Claude becomes the default model in Slack.

What this means for a GP

None of this is clinical advice, and none of it should change how you treat a patient tomorrow. It is why a busy doctor might still skim the letter. MHS is how a lab stops paying a specialist to hand-wire every new microscope — interesting if you sit on a pathology or research committee, not if you are trying to finish inbox. Whispering to agents is the same habit as dictating a letter: faster than typing once the microphone is good enough, still your name on the output. The makerskills trick — a named checklist that interviews you when you are stuck — is the bit you can steal without installing anything. The UK case is the one that will land in the waiting room as a headline. First patient, specialist centre, surgeon still in charge, larger trials next. If someone asks whether "AI did the operation," the honest answer is no: it watched, highlighted, and stayed on its own screen.

Sources

Plain-English notes on The Rundown AI · Dr Kotha · Gold Coast GP. Not medical advice. Source: The Rundown AI.