The AI bill you never see is now a creator problem, and Unprompt is the first tool I’ve seen try to price it
If you publish across five platforms a week, you are already running an AI operation whether you call it that or not. Captions in Claude, thumbnail variants in Codex, a Cursor session to wire up a scheduling script, a ChatGPT pass to repurpose a YouTube script into eight Threads posts. Every one of those keystrokes burns tokens, and every token burns electricity and water somewhere you will never visit. Until now, the only people who felt that cost were the ones paying API bills. The rest of us got a flat subscription and a vague sense of guilt. Unprompt is a small Mac and iPhone companion that tries to put a dollar figure and a water figure on that invisible usage — and for creators who treat AI as a production line, that framing is worth more than the app itself.
What Unprompt actually is, minus the launch-day gloss
The maker, Joëlle, builds three apps solo and says she had no idea what her AI tokens added up to — “in dollars, or in water.” So she built two things. On the Mac, a free companion app reads the local logs of Claude Code, Codex, and Cursor, converts token counts into dollars and liters, and never reads conversation content. On the iPhone, Unprompt reads Screen Time to count how often you open your AI apps, with work hours separated from personal ones. It blocks nothing. Today’s tokens and cost are free; the 30-day card with the water number sits behind Premium. There’s no account to connect, and totals sync through your own iCloud. Pricing for Premium is not disclosed in the launch copy.
The dollar number is explicitly not your subscription bill. It’s what your tokens would have cost at API prices — a counterfactual, and the maker says so plainly. The water number comes from electricity at roughly 4.9 liters per kWh, attributed to Berkeley Lab (2024). Her own last 30 days: 272 million tokens.
That 272M figure is the part I keep coming back to. If you’re a solo creator running an AI-assisted content pipeline, that number is probably closer to your reality than you’d like to admit.
The problem isn’t the money — it’s the missing feedback loop
Here’s the operational truth: creators optimize what they can see. We obsess over watch time, saves, sends, and engagement rate because the dashboards show them to us in real time. We do not optimize token spend, because nothing in our stack surfaces it. Buffer, Hootsuite, Later, and Metricool all tell you what your posts did. None of them tell you what generating those posts cost. That’s a real gap, and it’s not a small one once AI moves from “nice assist” to “the thing that writes the first draft of everything.”
Why this matters more to a TikTok creator than a LinkedIn one
Volume. A LinkedIn operator might ship three posts a week and call it a content strategy. A TikTok or Reels creator ships three hooks before lunch, tests them, kills two, and reships. Every iteration is a generation. Multiply that across a repurposing workflow — one long-form YouTube video chopped into nine verticals, each with its own caption set, thumbnail text, and alt copy — and the token curve stops being linear. It compounds. If you’re running that pipeline with CapCut for edits and an LLM for copy, you are the exact user Unprompt was built for, even if you’ve never thought of yourself as someone with an “AI spend.”
The friends-comparison feature is the sleeper hit
You can add friends and compare your month. In the comments, a user asked whether anything leaves the machine for that comparison — a fair question given the whole pitch is local-first. The maker’s answer: friends are opt-in, sync goes through Apple’s iCloud, and by default friends see your trend, not your exact numbers. Sharing exact figures is a separate choice. I’d bet that default is doing a lot of work. Social comparison is the single most reliable behavior-change lever in consumer apps — it’s why Strava segments and Duolingo streaks work — and a monthly token trend you can see next to a peer’s is a subtler, stickier motivator than any dashboard chart.
Where the math holds up, and where it’s honestly a vibe
This is the section I’d want if I were evaluating whether to trust the numbers. A commenter, Gal Dayan, made the sharpest critique on the thread: a single fixed average like 4.9 L/kWh hides enormous variance. Water use swings by datacenter, cooling design, and season. A request served by a free-air-cooled facility in a cool climate could be near zero; one served by a water-cooled facility at summer peak could be far above average — and the app has no way to know which one actually served you. His framing: the dollar figure is a real lower bound, the water figure is “more like a vibe.”
The maker’s rebuttal is worth quoting because it changes the picture. She clarified that the 4.9 L/kWh figure is mostly not cooling — it’s 0.36 L on-site plus 4.52 L for generating the electricity — so free-air cooling only removes a small slice. She conceded the grid mix varies and that energy-per-token is itself an estimate, which is why the app presents water as an order of magnitude, not a measurement. The FAQ walks through both steps and sources.
My take: that’s the right call, and it’s rare to see a maker defend a number by lowering its claimed precision. The dollar figure is the trustworthy one — it’s arithmetic against published API rates. The water figure is directional. If you screenshot the liters and put it in a client deck as fact, that’s on you, not the app.
Where the math breaks
Three places I’d flag before you build a workflow on this:
- Local logs only cover three tools. Claude Code, Codex, Cursor. If your stack runs through ChatGPT web, Gemini, Midjourney, or ElevenLabs for voiceover, those tokens aren’t in the Mac companion’s ledger. The iPhone side counts app opens via Screen Time, not tokens — so mobile usage is a frequency signal, not a cost signal. Those are two different measurements wearing the same dashboard.
- API-price framing flatters subscription users. If you’re on a flat $20 or $200 plan, your “cost” is fixed. The API number is what you’d pay if you were metered. Useful for understanding intensity, misleading if you read it as money leaving your account.
- No Windows, no Android, no browser extension — at least not mentioned in the launch. That cuts out a large chunk of the creator market, especially anyone editing on a Windows desktop.
None of these are dealbreakers. They’re scope boundaries, and the maker seems clear-eyed about them.
What creators and social teams should steal from this, regardless of whether you install it
The product is small. The pattern is big, and I think this is the real story for social media operators.
Audit your AI production line like you audit your posting cadence
Most creators can tell you their posting frequency to the day. Almost none can tell you how many AI generations it takes to produce one post. Run the exercise manually this week: pick one repurposing workflow — say, one YouTube video into verticals — and count every AI touch. Script rewrite, hook variants, caption sets, hashtag research, thumbnail copy, description, pinned comment. If the count surprises you, you’ve just found the same gap Unprompt is trying to close. You don’t need the app to do the audit; you need the habit the app is designed to create.
Local-first is a positioning lesson for every tool you build or buy
The strongest signal in that entire comment thread wasn’t the feature set — it was how fast the audience relaxed once they heard “reads local logs, no API keys, no proxy.” One commenter said they’d have been “way more hesitant if this needed my API keys or a proxy.” Another praised the “no login, no tracking” angle outright. That’s the market telling you something. If you’re evaluating any AI tool for your content stack, this is the question to lead with: what does it need access to, and where does my data go? Tools that answer “nothing” and “nowhere” are earning a trust premium right now, and I’d bet that premium grows as more creator workflows get automated.
The “order of magnitude, not a measurement” framing is a content strategy in itself
The maker’s own defense of the water number — “the goal isn’t precision, it’s to make an invisible cost visible and let you compare yourself to yourself over time” — is a clean articulation of how to talk about metrics you can’t fully verify. Creators do this constantly with analytics. You can’t know exactly why a Reel popped, but you can build a directional model and test against it. The discipline is the same: be honest about the confidence interval, then act on the trend anyway.
Where I think it falls short, and who shouldn’t bother
Let me be the skeptic for a paragraph. If you use AI lightly — a caption assist here, a grammar pass there — this app will show you a number so small it changes nothing. The behavior-change value only kicks in at volume. If you’re on Windows or Android, you’re out of scope today. If you want precise environmental accounting to cite publicly, this isn’t it, and the maker would agree — she calls water an order of magnitude. And if what you actually want is to reduce your subscription spend, this doesn’t do that; it just shows you intensity. It’s a mirror, not a brake.
The other open question, which the source doesn’t answer: how does the friends comparison handle wildly different workflows? A solo writer and a video team of five will have incomparable token profiles. Comparing your month to a friend’s is motivating until it’s meaningless. Not disclosed how the app frames that context, if at all.
And one more, since I write for operators who care about client work: there’s no mention of team or agency seats. If you run social for multiple brands, you’d want per-client attribution of AI cost. Not in the source, so I won’t assume it exists.
What I’d watch / test next
Concrete moves for this week, whether or not you install anything:
- Run a one-week token audit. Pick your heaviest AI-assisted workflow and count generations manually. If the number shocks you, you’ve validated the problem Unprompt is solving — and you’ll evaluate the app honestly instead of on launch-day hype.
- Check your stack’s local-log coverage. If your AI work happens mostly in ChatGPT web, Gemini, or mobile apps, the Mac companion’s Claude Code / Codex / Cursor coverage won’t capture it. Know your blind spots before you trust a total.
- Treat the water number as directional, the dollar number as arithmetic. Use the dollars for decisions. Use the liters for storytelling — carefully, with the “order of magnitude” caveat attached, especially in client-facing decks.
- Watch for Windows, Android, and team features. Those are the three things that would move this from a solo-creator curiosity to an agency-usable tool. If the maker ships any of them, I’d revisit.
- Ask the local-first question of every AI tool in your stack. Not just this one. It’s the right filter for 2025 and beyond.
The bigger story here isn’t Unprompt. It’s that AI usage is becoming a line item creators can no longer ignore — and the first tools to make it visible will shape how we think about it. Whether this particular app wins or not, the category it’s opening is one every social media operator should be paying attention to. I’ll be watching whether it stays a mirror or becomes a control panel.
