Thursday morning. I’m staring at my Anthropic bill wondering “how much have I actually spent today?” and there’s no easy answer.
No menubar widget. No quick glance. Just… log into a dashboard, find the billing page, squint at numbers that update whenever they feel like it.
So I told my AI agent to build one.
The 2-Hour Build
TokenMeter. A native Mac menubar app. SwiftUI, no Electron bloat — the whole thing is 394KB. Here’s what it does:
- Sits in your menu bar showing today’s spend
- Tracks Anthropic, OpenAI, Google AI, and xAI
- Budget alerts before you hit a surprise bill
- For Claude Max subscribers: shows messages remaining + reset timer
Concept to working app: about 2 hours. Concept to signed, notarized, Apple-approved DMG: same session.
My AI agent (running on OpenClaw) wrote every line of Swift. I don’t write code. I describe what I want, review what comes back, and push back when it’s wrong. The agent handles the build loop — write, compile, fix errors, iterate.
The Plot Twist
Here’s where it gets interesting. We wired up the real API calls and discovered something most people don’t know:
You can’t get your AI spending data with a regular API key.
Both Anthropic and OpenAI lock their billing/cost endpoints behind “Admin API keys.” These require an Organization account. Not a paid org — creating one is free — but you need to know to do it, find the right settings page, create the right type of key, and grant the right role.
We’d built the entire app, polished the design, set up the landing page, got Apple notarization… and then hit this wall during testing.
The app works perfectly. The onboarding just needs to guide users through a non-obvious setup step that every single cost-tracking tool in the market also requires (Vantage, CloudZero, all of them). We just didn’t know that going in.
What I Actually Learned
Test the data pipe before building the UI. 20 minutes of curl requests would have caught this before we wrote a single line of Swift. We’ve now added a “Feasibility Gate” to our product evaluation process — actually call the API with a real account before building anything.
AI agents are absurdly good at native app development. SwiftUI, entitlements, code signing, notarization, DMG packaging — the agent handled all of it. The compile-fix loop that would take a human developer hours of Stack Overflow happens in seconds.
The 394KB lesson. The finished app is 394KB. A typical Electron “menubar app” doing the same thing would be 150-300MB. Native matters for always-on utilities.
Dead projects aren’t dead content. Even if TokenMeter never sells a single copy, this post exists. The insight about Admin API keys helps anyone building in this space. The pattern of “test before you build” saves someone else the same mistake.
Current Status
TokenMeter is built, signed, notarized, and has a landing page. We’re testing it ourselves now (getting those Admin keys set up), then it goes on Gumroad for $5.
Will it make money? Honestly no idea. But the build cost me nothing except API tokens I was already paying for, and now I have a tool I actually use plus a story worth telling.
That’s the game. Build fast, learn something either way, share what you find.
Building in public @wayneb. This is project #3 in my hunt for $50k/month independent income.