What’s connected to my AI – and what isn’t

Accessed through the Claude desktop app, sometimes the plain terminal, and sometimes from a browser – the Code tab at claude.ai shows the same session that’s running on my Mac, so I can watch it and steer it from anywhere. This is where Claude works directly in my files – my whole system runs through it, daily. It reads my vault, runs my skills, and does the busywork while I direct.

Two lines worth being explicit about: the files stay on my machine. iCloud backs them up, but Claude works on the local copies – there’s no cloud drive in the loop. And my account is set to keep my conversations out of model training.

A small, stable set of foundational files uploaded to a project in the iPhone app, so the phone knows my voice and my work. A snapshot, not a live connection – good enough for thinking out loud when I’m out and about. When a chat produces something worth keeping, I have it write a markdown file, and I carry that into the vault myself. By hand, on purpose.

Two things. One: I canceled the subscription but kept the export – about 2,500 conversations from 2023 to 2026, searchable from my machine. Not as useful as I expected, but it’s mine to search when I remember something worth pulling out of the archive. Two: I have access to an enterprise account through a client – and the ground rule above keeps client work off this board. I don’t use it often.

Since June 2026: a second, unrelated agent pointed at the same files, so I can get a second perspective and my system keeps running if something goes wrong on Claude’s side. Staying AI-agnostic is nice to have – I’d be overstating it to call it a requirement while I’m not actively doing it. Since then I split the vault into a main vault with satellites, and the new setup hasn’t been through it yet. On the list to re-test.

Gemini and I have two histories. One: the web version, where I upload a video and ask for a shot-by-shot read – what happens on screen, when, and why it works – while I teach myself how video actually gets made. That still happens now and then. Two: its API transcribed the footage in my early rough-cut pipeline, supplying clean text while another service carried the timestamps. That pipeline is retired; the key stays in the toolbox in case a job ever shows up again. Either way: nothing private goes in, and nothing where accuracy matters comes out unchecked.

When I need the lay of the land on a topic, Claude prompts Perplexity directly and gets a report back. Then it follows the citations, reads the primary sources, and builds its own synthesis from what they actually say. The report is a map, not the territory – we don’t take it as fact.

Searches the actual papers – PubMed is the free index of biomedical research – for the writing work where being wrong has consequences. Public data, low risk, high value: the easiest yes in this catalog.

This replaces an SEO tool subscription I would otherwise be paying for, and does the same job: real search volumes, real results pages, related keywords, the questions people also ask – even where the AI answer engines mention a brand. The point is the pricing: I pay cents when I need data, through an API, instead of a monthly fee whether I use it or not. Everything I’ve pulled to date has cost about $13 total.

My task and project system – Thomas Frank’s Ultimate Brain, with Creator’s Companion alongside. Areas of my life, notes on topics, projects and tasks. Claude reads from it and writes to it – the vault routes to-dos straight in – and it follows me to my phone. Content planning used to live here too, but it never matched the way my brain works; that moved next door.

Video and content planning live here, in a base Claude reads and writes through the API. This format matches how I want to see things organized: what’s planned, what’s filmed, what’s cut, what’s scheduled, what’s live – and the point behind each of them.

Both newsletters – this site’s and the hobby blog’s – run on Kit, and Claude reaches each through its own API key, stored where it can be used but never read. It stages drafts; sending is mine alone. Nothing mails itself.

Connected July 2026 so Claude can plan my working blocks – two hours here, three there – around what I need to get done that particular week. The mechanics are boring, and that’s the point: no login, no account access granted. Claude reads the calendar app right on my Mac, and that syncs to my phone.

Drives a real browser for the things APIs and other interfaces can’t reach – what a search results page actually looks like, what a competitor actually shipped, filling out forms, setting things up in web dashboards, thinking through the privacy settings on a social account with me. Anything that has to be set up by hand in a browser.

And one line about logins: this browser doesn’t sign me in anywhere by itself. My logins live in a password manager, so Claude can only reach what I’ve logged into myself, in that browser, while it works.

The hobby blog and this site, both running WordPress on a German web host – my own installs, not WordPress.com. The AI publishes through accounts I created specifically for it, scoped to the least power that does the job; a guard refuses to overwrite anything I edited by hand, and nothing goes live without my word. The hobby blog came first – it’s where every workflow on this page made its mistakes before the mistakes had an audience. My freelance-business site is next on the list.

One of the first automations I ever set up, long before I even started using Claude Code. I record a voice memo in the RecUp app on my phone; it lands in a watched Dropbox folder, a Pipedream workflow transcribes and summarizes it, and the note – along with the full transcription – ends up in my vault, where Claude can read it and do something with it. Adapted from Thomas Frank’s template – I changed one step so it goes to my vault instead of to Notion. It’s been running ever since, and it’s the best automation I ever built.

This one simply is Thomas Frank’s automation, unchanged: an iPhone shortcut sends spoken tasks through a Pipedream workflow into my Notion board, with due dates and project assignments filled in. I don’t use it as often as I should.

An online transcription service, and part of the video-editing pipeline I built first: raw footage goes to its API, and back comes a word-level transcript where every word carries its timestamp. That’s what lets a system cut video by editing text – the words drive the knife. Every video so far has started here, whatever happens downstream.

The last stop before upload – and honestly, this one is for me. An earlier pipeline had a studio-sound pass I loved, and nothing since has matched it; Auphonic comes closest. One API call, pay-per-credit, the video stream untouched – and my voice comes back sounding like something I can stand to hear on repeat while I edit and review.

Since July 2026: the editing pipeline I’m testing as a successor, after running into the first one’s limits. To be fair, no video has gone end to end through it yet – the latest was a hybrid, the cut made in my own pipeline and Descript handling scenes and visuals, with Claude driving it through the API and me finishing by hand in the app. Still working it out.

Based on an idea from Andrej Karpathy: one question fanned out to four frontier models from different companies; each reviews the others’ answers anonymously, and a fifth synthesizes. Runs pay-per-use through the OpenRouter API. It was fun to build and it’s fun to watch. I don’t use it often – it shines when you already know your answer and want to see whether an anonymous jury agrees.

Connected for social publishing, then reality arrived: its upload cap sits well below the size of a real YouTube video, so long-form goes up by hand and this connection mostly sits. Whether it stays for the short stuff is an open call. It’s in the catalog because the catalog records decisions, including the ones still pending.

My whole system is markdown files, and Obsidian is where I read and edit them as a human. A local app over local files – when the AI is down, this is how everything keeps working. Its Web Clipper in my Chrome is the one piece that touches the web: it clips articles and pages into the vault for later reading.

Renders my video graphics – title cards, overlays, animations – from code instead of a drag-and-drop editor. The AI writes the code; I judge the frames. I’m still not sure I love it: the Descript trial above is me testing the alternative.

Every cut, conversion, and audio check in my own pipeline runs through it. If the Descript route wins the trial, this one gets quieter days. We’ll see.

The open-source Whisper speech-to-text, running on my own machine. When I download a video to learn from it, this is what turns the audio into text I can read and search – nothing leaves the Mac.

An open-source downloader: hand it a video URL and the clip lands on my disk as a file. The Web Clipper handles articles; this handles video – the things worth learning from that have no transcript to clip. Downloaded clip in, whisper.cpp transcript out, and I can study it properly.

Turns one document format into another – markdown to Word and back, mostly – without opening an app.

That’s really its name, and it does exactly what it says: takes a scanned PDF and adds a text layer, so paper documents become searchable files.

Paid, not open source – and worth it for one feature: I can teach it my vocabulary, so the words I actually use come out spelled right the first time. Claude can decode a mangled transcript; an email or a document can’t wait for that. And it runs right on the machine – what I say stays here.

Also local, also Whisper-based. VoiceInk took over the day I wanted to teach my dictation tool my own vocabulary.

Email is where my clients message me in confidence about their projects, and that makes it too risky to connect. I’m in Germany, many of my clients are in Europe, and GDPR is part of every decision like this – I will not put client email conversations on an LLM provider’s servers. So my inbox stays mine alone, on purpose; that’s why email is the one tool missing from the homepage map. Maybe other inboxes get connected at some point. Not yet.

Putting my vault on GitHub would buy real convenience – Claude Code could work on it from any browser, from anywhere. But my machine holds client work, and under GDPR I’m responsible for every place that data travels; a cloud repo is one more processor I’d have to vet, justify, and contract with. Easier and safer to skip it: version history happens in git on my own machine, and the files stay where I can point at them.

The only things that land in Google Drive are my blogs’ backups, pushed there by a WordPress plugin under its own accounts. Claude’s not connected. No plans to add it.

There’s no team channel in my working day – I’m one person, and the people I work with reach me by email.

My automations run on my own machine, where I can read every step. A cloud middleman would add accounts, tokens, and someone else’s privacy policy to jobs a local folder and a script already do.