v0.4.0: An iPhone companion

What it actually does

It opens on a question rather than a screen full of controls: what your business is, and what you send your accountant. One sentence, typed or spoken. It reads back what it understood, you fix anything it got wrong, and that is where you pick the month — it is the only place a period is ever chosen.

Then Clerk.AI opens a browser and starts working through it: signing in where it can, downloading the statements, then going after the proof behind each line. Invoices in your mail, receipts in the billing portals you use, the photo you took of a receipt in a shop, which it reads with the OCR built into macOS.

What you watch while that happens is the email your accountant is going to get. It is on screen from the first minute, mostly grey, and each line goes solid as the thing behind it turns up. You do not have to learn what the app is doing — you can see how much of the letter is real yet.

It works in runs of about an hour, in the background, and you are not meant to sit and watch it. A month usually takes a few of those. How long the whole thing takes on the calendar is mostly down to how quickly you answer it, because when it needs you it stops and waits.

It needs you a handful of times per close, and most of those are the same request: get me past a login. It gets itself into most places on its own, and it can even fish a one-time code out of your inbox to finish the job, but anything that wants to prove you are you — an app to tap, a code to scan, a passkey — is you, on your machine, and it should stay that way. That is not only banks: brokers, exchanges and any supplier portal sitting behind bank identification hit the same wall. The rest are the judgement calls it cannot make, like whether a particular lunch was a client expense or which project a flight belonged to.

Photos are the one place it deliberately over-collects. Anything on your phone that looks like a receipt gets offered to you as a grid, each one with what it managed to read off it, and you include or drop them one at a time or accept the rest in a single press. Dropping is remembered, so the same photo of somebody's dog is never put in front of you twice.

At the end you get a flat folder named the way your accountant likes, and the email in whatever language they read, with every attachment in order. You look it over and you send it. The app will not mark your books closed for you, and it will not send anything on your behalf.

A period that has been sent stays open to what happens next, which is the part of the handover that used to fall on the floor. Your accountant writes back asking for the one thing that was missing; you paste that in, and it goes and fetches it and remembers for next time, so the same gap does not come round again in a month.

Answering its questions from your phone

That is the new part in 0.4.0, and it matters more than it sounds. In the close I have logs for, the longest single stretch was fifty minutes of the app sitting there waiting for me to come back and say one sentence. Every one of those interruptions used to mean returning to the Mac, which is a silly reason to have to sit down, so there is an iPhone app.

It opens on what the close is doing and when it last moved, which is the thing you want to know when you pick up your phone: whether it is working, waiting on you, or stuck. When there is a question, you hold the button, say the answer out loud, and put the phone back in your pocket. The recording goes over to your Mac, gets transcribed there by Whisper on your own hardware, and the close carries on from where it stopped. There is nothing to wait around for.

Those logins are the exception, because they genuinely need you at the machine. When a close hits one, the phone tells you it needs your Mac, says nothing else is blocked, and offers you whatever other questions it can get out of the way in the meantime.

Sync goes through your own iCloud, between your Mac and your phone. The companion is meant to sit behind a licence, at €5 a month or more plus VAT, and while it is in beta everyone gets it anyway. That will change at some point. If you are already paying by then it will not change for you, because what your amount includes is fixed on the day you subscribe. Everything else described here is the free app and stays free.

One thing to be straight about while this is fresh: the iPhone app is not in the App Store yet, and it is going in as soon as we can get it through review. Until then it is on TestFlight, and anyone who wants it can have it. Ask and we will send you an invitation. Nothing on the Mac depends on it.

Try it on a company that doesn't exist

The reasonable thing to be nervous about is letting software drive a browser that is logged into your bank. I would be.

So when your first close starts, it offers you a made-up month to watch while you wait. That opens in its own window on a made-up company with made-up receipts, keeps its own storage, and never touches your real books. Turning it down costs nothing and it does not ask again. One thing to know: its questions reach your phone alongside real ones, because every open vault syncs through your own iCloud, so a demo month will put fictional questions in your inbox. Watching it do a month for somebody fictional answers that question better than anything I can write on this page.

What you need

A Mac, an accountant you already send things to, and your own Anthropic subscription, because the app drives Claude Code and you bring your own account for that. We do not resell AI usage. Setting Claude Code up involves a terminal.

It drives a separate Chrome profile that you sign into yourself, so your everyday tabs and logins are not part of it. That is also why it works with banks and portals nobody has built an integration for: it is a real browser doing what you would do.

Setting it up

The close itself works. I have been closing my own books this way for nine months, the last two with the app and the seven before that driving Claude Code by hand, one prompt at a time, before the app existed. Setting it up is the part that still costs you something real, and what 0.4.0 changed about it is the order rather than the contents.

It is three screens. The first is the question about your business, and nothing is asked of you before you have answered it: no terminal, no permissions, no account. Typing needs no permission at all, so if you type you are never interrupted by a system dialog before you have said a word, and the microphone is asked for the first time your hand reaches for the mic, which is the only moment that request makes any sense.

The second screen is the technical part, and it arrives only because you pressed go. It wants one line pasted into Terminal, which installs the engine, and it wants to know who is paying for the AI. Your answer stays quoted at the top of it the whole time. There is no Continue on that screen: the button is deliberately dead and says what it is still waiting for, and when the last check goes green the screen becomes the read-back on its own. The terminal line is where people get stuck, and I would rather say so here than let you meet it the night before a deadline.

Before anything acts on your behalf, that screen asks you to accept what you are letting it do, and the other button on that card quits the app. Then the third screen reads back what it understood, you fix anything it misheard, and the letter takes the window.

Those steps now come in the order they matter, which is the other thing this release changed. You answer the question about your business first, before anything technical is mentioned — so by the time it asks you to open a terminal you know what it is for, and your answer is already sitting there waiting. The app went from something you had to learn your way around to something with one thing on screen: the letter, filling in.