We Charged AI Agents $0.12 Real Money. Here's the Exact Stack.

Not a typo. Not a MRR flex. Twelve cents — from two AI agents we have never spoken to, who found our API on their own and paid per call in crypto. This is the exact stack, including what broke.

total collected · 5 paid calls · 2 external agents · $0 spent on distribution

The stack (all running on one $10 VPS)

• FastAPI — the parser endpoint (Excalidraw → markdown/JSON)
• x402 middleware — unpaid requests get an HTTP 402 with a machine-readable price; paid retries carry a signed proof we verify before doing work
• Per-call tiers — free 5/day, $0.005 basic, $0.05 premium (we doubled this price with zero churn)
• Docker on Oracle Cloud free tier — yes, the whole revenue engine is on the free tier

The part that actually drove revenue

Not SEO. Not ads. Machine-readable documentation. We publish /llms.txt, /agents.json, and a full OpenAPI spec. Agents crawling for capabilities read those files, see the price and the 402 flow, and just… call. Two of them did. That's the entire acquisition funnel.

What broke along the way

• Our webhook silently died — Vercel freezes fire-and-forget fetches after the response. Leads vanished with a 200 OK. Fixed by awaiting the fetch.
• Google Workspace blocks SMTP app passwords — our entire nurture-email plan died until we switched to an API sender.
• We shipped a price change and forgot the container caches constants at import. Restart fixed it; the docs now say so.

What's next

Registry listings so 100× more agents find us, productized feeds ($99–499/mo) for humans, and a public revenue counter on the homepage that updates live — because the counter being small and real is the whole point. Watch it at nanoempireai.com.

If you have a Python function, you can put it behind the same rails with the tollbooth SDK — free, MIT.