DEVONthink 4.3 and a Privacy-First MCP

DEVONthink 4.3 shipped last week, and the headline for me is a single feature. DEVONthink now has its own MCP server.

If you’ve been following along with the Robot Assistant work, you know why that matters. MCP is the connector standard that lets an assistant like Claude reach into an app and actually do something.

Until now, getting Claude to work with your DEVONthink databases meant exporting files or building your own bridge. Version 4.3 closes that gap. The new server exposes close to 60 commands, so Claude can search, read, and organize inside your databases directly.

What I appreciate is how carefully DEVONtechnologies built it. They designed the server with privacy at the center. You control both local and remote access. It redacts sensitive information before anything leaves your Mac. It respects items you’ve already marked as excluded from AI. And it blocks direct filesystem access to the database internals, so the assistant works through DEVONthink instead of around it.

There’s a new Exclude from Chat & MCP option too. Some databases simply shouldn’t touch an AI, and now you can flag a whole database at once. DEVONthink does this automatically for encrypted and revision-proof databases, which is the right default.

The chat assistant inside DEVONthink got the same attention. It now redacts a wider range of sensitive data, including phone numbers, credit card numbers, authentication tokens, and anything you’ve labeled as a secret, before that text reaches a language model. That protection carries through to the AI smart actions and the summarize and transform features.

Connecting an AI to your knowledge base is a real decision. A lot of MCP servers shipped over the past year without much thought about security, and the gap showed. DEVONthink clearly thought about it.

The rest of 4.3 is a solid update on its own. There’s a new Markdown engine that replaces MultiMarkdown and adds callouts, citations, captioned tables, and better CriticMarkup. Markdown documents now handle smart quotes and dashes natively. There are new desktop widgets for macOS Sequoia, a handful of improvements to the web interface for DEVONthink Server users, and a rebuilt PDF viewer with a table of contents, page thumbnails, and full-text search.

DEVONthink 4.3 is a free update for 4.x users and needs macOS Ventura or later. If you keep your files and research in DEVONthink and you’ve been curious about putting an assistant to work on them, this is the update that makes it possible. After spending a week with the update, I’m finding myself loading even more in DEVONthink.

Cotypist Brings Smart Autocomplete to Every Mac App

Daniel Gräfe launched Cotypist to the public yesterday, and it’s worth checking out.

Cotypist is smart autocomplete for the Mac. Not autocomplete in one editor. Autocomplete everywhere you write. Mail, notes, documents, messages, the prompt box of whatever AI tool you’re using.

You start typing, a suggestion appears, you press Tab, and you keep going. It learns your voice as you write, and it carries that voice across every app instead of starting fresh in each one.

Daniel made this app to scratch his own itch, and in my experience, those are always the best apps. The part that matters most to me is where the work happens. Cotypist runs on-device. What you type is some of the most personal text you generate. Private notes, half-formed ideas, drafts you would never want sitting in someone else’s training data. Keeping all of that on your own Mac is the right call.

It’s been improving fast. The recent 0.22 release added inline autocorrect, a Cotypist Labs section for experimental features like mid-line completions, British English, and right-to-left support for Arabic and Hebrew.

Cotypist has a three-tier model. Free gives you 100 accepted completions a day, which is plenty for light typing. Plus is $6 a month or $72 a year for unlimited completions and better models. Pro is $9 a month or $108 a year and adds the full model catalog, per-app instructions, and early access to Labs features.

Suggestion-based writing isn’t for everyone. Some writers find a completion popping up mid-sentence more distracting than helpful. The way to know is to try it, and the free tier means trying it costs nothing. The app is spooky-good. I’m in.

Bartender 6

Bartender 6 is out, and if you have a MacBook with a notch, the headline feature is Top Shelf.

The idea: that camera notch at the top of your MacBook display is real estate that Apple has never done much with. Bartender Pro turns it into a useful surface for widgets, media controls, files, and clipboard history. Hover over the notch area and a panel expands. Move away and it disappears. It’s a clean solution to a space problem Apple created and then mostly ignored.

Top Shelf is part of the new Bartender Pro tier. Whether that’s worth it depends on how often you find yourself wishing something useful was up there.

The rest of Bartender 6 does what Bartender has always done, now with full macOS Tahoe support and Liquid Glass compatibility. If you owned a previous version, the upgrade is free.

It’s good to see version 6 in the wild.

Get Bartender 6

Mac Power Users 850: Overcast Transcripts and 48 Mac minis with Marco Arment

On this episode of Mac Power Users, Stephen and I sit down with Marco Arment to talk about what happens when a podcast app developer accidentally builds a data center. Overcast now transcribes podcasts on demand, and the hardware behind that feature is 48 Mac minis running in a rack. Marco walks us through how the system works, how he manages it all remotely, and what he’d do differently if he started from scratch.

We also catch up on the rest of Marco’s life, including the three-year-old MacBook Pro he refuses to upgrade and the restaurant he now helps run. There’s a good stretch on using AI in an actual development workflow and where Marco hopes Apple takes things for developers.

Episode Links

This episode of Mac Power Users is sponsored by:

  • NetSuite: The leading integrated cloud business software suite.
  • Mercury Weather: Forecasts, beautifully done. Download now for free.
  • 1Password: Never forget a password again.

The Silent Film Era of AI

The first movies were shot from a theater seat. The camera sat where the audience would have sat. The actors moved across the frame like they were on a stage, and the shot didn’t change for the whole picture. You were watching a play that someone had happened to have filmed.

It took years before folks had the idea to put the camera on the stage. That single move turned a recorded play into cinema.

I think we’re in the silent film era of AI right now.

Most of us approach a problem the way we did before any of this existed. We pick up the same tools we used yesterday and ask the model to make us a little faster. That’s the camera in the theater seat. Useful, but limited. Not really cinema.

The most interesting moves are happening in the corners, where someone forgets the old rules and asks a different question. Rather than use the robot to drive a particular piece of software, you have it develop a system that meets your needs more specifically than the software ever could.

Those of us interested in automation have spent years building Rube Goldberg machines that kind of get it done. Hazel is watching a folder, a Keyboard Maestro macro firing in response. The whole thing held together with twine and chewing gum. AI lets you wipe the slate clean and replace complicated workflows with simple ones that happen in the background.

Instead of stringing five apps together to file your receipts, you describe what you want, and the robot watches your downloads and puts them where they belong. Instead of a tag system across three note apps to find old research, you ask the robot a question, and it reads the folders and answers. Instead of an inbox with twelve filter rules and half a dozen labels, you tell the robot what you care about, and the rest happens before you sit down.

None of those systems existed two years ago. All of them came from someone asking what the new way looks like, rather than how to speed up the old way.

That’s the move I’m trying to make myself, and it’s harder than it sounds. The pull to stay in the old frame is strong. My instinct is still to use the robot to do the same thing I would have done, just faster. When I catch myself doing that, I try to back up and ask the other question. What does this problem look like if I forget how I would have solved it last year?

It leaves me with one question for you. What problem are you still solving the old way because you haven’t asked what the new way would look like?

FotoMagico, Revisited


FotoMagico just hit 6.9. Boinx has been quietly shipping updates all year, which is a good excuse to revisit an app I suspect most of you have forgotten about.

For years, I was the designated slideshow person in my family. Weddings, memorials, graduation parties. Someone would hand me a folder full of photos and say “can you put something together for Saturday?” and I would open FotoMagico and get to work. The app made it easy to do the job well: real transitions, precise timing, Ken Burns effects that actually looked intentional, and music sync that didn’t fight you.

At some point I drifted away. The Photos app covers the basics now, and most occasions don’t demand more than that. But “the basics” is a different thing from standing in front of 150 people at a wedding reception with a twelve-minute show on a MacBook.

One piece of practical wisdom I always passed on: tell the family you need the photos a week out, and tell them the show runs about ten minutes. Eight is better. Nobody in the room actually wants a fifteen-minute slideshow. They think they do.

I’m delighted to see version 6 in the wild. The app is still actively developed, still well-designed, and still the right answer when the stakes are high enough to matter.

If you already own FotoMagico 6, the 6.9 update is free. New buyers can grab it on the Boinx developer page.

If there’s a family event coming up and someone’s already started texting you about “the photos,” now you know what to download.

The Productivity Apps That Don’t Speak Agent Will Lose

Spark Mail shipped a Mac CLI and agent skill hooks last week. Read-only email access is free. Write actions (send, move, archive) are behind the Pro subscription. Readdle also published open-source automation recipes and persona hooks for anyone building on top of it.

I want to set the specific features aside for a moment and talk about what this move represents.

The people who use productivity apps in 2026 are asking a different question than they were three years ago. It used to be: does this app do the thing well? Now it’s also: can my robot assistant work with this app?

That’s a real shift, and it’s not going away. Spark’s CLI is a direct acknowledgment that the app’s users are increasingly expecting their tools to “speak agent”. You can frame it as a niche developer feature if you want. I think that’s the wrong frame.

The users who are most engaged, most willing to pay, and most likely to recommend software to other people are exactly the ones building personal automation pipelines and wanting their apps in the loop. Readdle didn’t have to publish open-source recipes. They did it because they understand who their users are becoming.

Productivity app developers who sit this out are taking a risk. The tools that figure this out early will have an advantage. The ones that don’t will find their users migrating toward whatever does.

Mike Schmitz’s New Ideal Week Builder

Mike Schmitz shipped a free web tool for designing an ideal week template, and it is pretty clever.

The interface is a calendar grid. You drag blocks for your work hours, deep work, family time, exercise, whatever modes matter to you. You color them. You theme the days at the top. Then you export the result as a PDF, an image, an ICS file, or a Markdown document.

The export options are what set it apart. The ICS file means you can drop your ideal week into Apple Calendar or Google Calendar as a toggleable planning layer. Turn it on when you are sketching out a week. Turn it off when you are looking at the actual schedule. The Markdown export means anyone working in Obsidian can paste the template straight into a personal retreat note, a quarterly planning doc, or a roles-and-goals review.

I sat with the tool on the air while Mike walked through his on the latest episode of Focused. It slotted right into the way I already think about weekly planning. Designing an ideal week is one of the most useful things you can do for your time. Most people never do it because the friction is high. Mike’s tool brings the friction down to almost zero.

It is free. Give it a try.

Still on Kagi, Still Happy

John Gruber wrote about Kagi again this week, which is a good excuse for me to update my own take.

I’ve been a paying subscriber for a couple of years now, and I have no plans to stop. My search results are better. They come from sources I actually trust. And I’m not being tracked, profiled, or served ads based on whatever I looked up yesterday.

The subscription is $10 a month. I don’t have a Starbucks habit, so I don’t find that hard to justify. Every search I do is one less transaction with an advertising machine. That’s worth something to me.

The knock on Kagi has always been that AI search will make dedicated search engines obsolete. I don’t think that’s right, at least not yet. AI gives me a synthesized answer. Kagi gives me a set of sources I can actually go read. Those are different things, and I still need both. For “what’s a good PDF editor for Mac?” I want a synthesis. For “what’s actually happening with Apple’s App Store case in Europe?” I want links to primary sources, not a summary.

Gruber notes that he occasionally uses Kagi’s Bangs feature (!) to fall back to Google. That’s the right way to use it. Kagi first. Google as a last resort.

If you haven’t tried it, the first 100 searches are free. You’ll probably notice the difference within the first ten.