I burned $13.5k of AI tokens (at list price) building a Mac app. Revenue so far: $0.

I burned $13.5k of AI tokens (at list price) building a Mac app. Revenue so far: $0.

Numbers first, because that's what I'd want from this post:

• Revenue: $0. Users: roughly zero. A handful of downloads so far.
• Build time: 31 active days across 7 weeks, evenings around a day job. I fire off a task when I finish work, then knuckle down from about 9pm to 2am. Up again at 7:30.
• Lines of code I've written myself: 0.
• Claude tokens: 61 million output tokens (41M of that on Opus), plus 15.7 billion cache reads. At API list prices that's about $13,500 of usage. I paid for it through a $200/month Max subscription, plus Codex and a SuperGrok plan I've since run dry.
• The app is live: a native Mac dictation app, $19 one-time after a free window, everything on-device.

So: a working product, a five-figure token bill at sticker price for a $19 app, and nobody using it. Here's the part worth arguing about.

I don't write the code. I run the team.

Every change ships through a team of AI agents: one builds, one reviews, one runs QA gates. Nothing I type ends up in the repo — my job is specifying what to build, reading diffs hard (much more in the early weeks than now), and being the one gate that can't be automated: I live-test the actual build with my actual voice before anything counts as done.

The review step is the part I'd defend at midnight: I make models from different labs argue. GPT-5.5 and Claude adversarially review the same diff, and they are ruthless with each other in a way neither is with its own work. Watching two frontier models fight over whether an audio buffer is being handled correctly is the closest thing this project has to a senior engineering culture.

What went wrong (a partial list)

The agents fought each other before the reviewers did. Early on I had multiple agents working the same checkout in parallel. Each one saw the others' edits as unexplained changes — and helpfully reverted them. Work just vanished. The fix was embarrassingly simple: tell each agent the others exist. Later I moved parallel feature work into separate worktrees entirely. I use worktrees every day at my day job; not starting there was my mistake, not theirs.

Days lost to a language argument. I wanted the HUD's visuals written in Rust and compiled to WASM. The agents kept "helpfully" doing it in JavaScript. I'd specify Rust; they'd bounce back to JS. It took multiple days of increasingly explicit instructions to make it stick. Agents don't share your architectural taste — they have to be told, in writing, repeatedly.

$40 of Kimi K3 for zero merged changes. New model launched, I gave it a shot, and the QA agent spent the credits driving browser automation. I got an audit document and nothing shippable. Not every model slots into an existing process.

The all-nighter the deprecation caused. The day before a model I leaned on was scheduled to lose access, I burned through 80% of a week's subscription quota in 24 hours getting everything I could out of it. Building on other people's models means their calendar is your calendar.

What it's actually cost

• $200/month Claude Max, plus the Codex and SuperGrok subscriptions
• $40 in Kimi tokens (see above)
• $50 for a PeerPush listing — which, I'll admit, worked: #2 Product of the Day yesterday, 20 upvotes
• ~5 hours most weeknights, and one week lost to Google quarantining my site because I'd skipped a notarization step (story for another post)

And what that bought: a shipped v0.6.0, ten blog articles, eleven free browser tools as top-of-funnel — and 210 visits and 560 page views in the last 30 days, a handful of downloads, zero sales. The product exists. The business doesn't, yet.

What I want from you

If you write all day on a Mac, I want you as user #1 through #10 — and I'll make it worth the detour: the first ten people who try Clabrate and give me honest feedback on the actual app get the $19 license free. Honest means what broke, what felt slow, what you'd never use — not encouragement. There's a free window to test in: hold a key, talk, clean text lands in whatever app you're in — or hand a task to the assistant and let it drive the Mac. Nothing leaves your machine. What you hate, I'll fix, with receipts, in next month's numbers post. I'll keep publishing everything: traffic, downloads, revenue, token spend.

And the question I actually want to argue about: I ship code I never wrote, reviewed by models arguing with each other, gated only by me using the build every day. Where's your line? Would you ship this way — and if not, what's the check you wouldn't hand over?