How an animal sanctuary owner in rural Japan — zero engineering background, 28 cats and dogs — built a production API platform using only AI agents.
30+ API integrations. Token economy. Docker deployment. Multi-language support.
Zero hired developers. Every lesson learned the hard way.
Where this was built: the cat cafeteria at Washin Village, rural Japan.
This is a real story. Every number is real. Every failure happened. Every lesson cost hours.
I run an animal sanctuary in rural Japan. 28 cats and dogs. My background is in business, not code. I'd never written a line of JavaScript before 2025.
In 7 months, I built a production API platform — live, serving real users, handling real money — using AI coding agents as my entire engineering team.
This document is exactly how.
| Metric | Value |
|---|---|
| APIs integrated | 30+ (translation, vision, search, social media) |
| Lines of code | ~50,000 |
| Engineers hired | 0 |
| Monthly AI subscription cost | ~$500 |
| Time to build | 7 months |
| Production uptime | 99.5%+ |
| Skills extracted | 112 reusable patterns |
| Bugs survived | 200+ (each one became a skill) |
| 3am outages | More than I'd like to admit |
Month 1 — "How hard can it be?"
I wanted to build an API marketplace. Something that would let AI agents access various services through a single platform. I'd seen enough tech Twitter to know the tools existed.
My plan was simple:
- Tell Claude what I want
- It writes the code
- I deploy it
- Profit
Here's what actually happened:
Day 1: "Build me an API server"
→ Claude writes beautiful code
→ I have no idea how to run it
Day 2: "How do I run this?"
→ Learn about npm, bun, node
→ Pick bun because Claude said it's fast
Day 3: "It works on my laptop!"
→ Try to deploy
→ What's a VPS? What's Docker? What's SSH?
Day 7: First endpoint goes live
→ It crashes under first real request
→ The error message means nothing to me
Lesson learned: AI can write code. But code is only 20% of shipping software. The other 80% is everything nobody tells you about — deployment, monitoring, error handling, security, databases, DNS, SSL, Docker networking...
Month 3-4 — "One agent isn't enough"
Claude Code is incredible at thinking and planning. But it can't see images. It can't search the web in real-time. And it gets expensive if you ask it to do bulk mechanical work.
So I built a team:
┌─────────────────────────────────────────────┐
│ Claude Code (Boss) │
│ Plans, architects, writes complex logic │
└──────────┬────────────────────┬──────────────┘
│ │
┌──────▼──────┐ ┌──────▼──────┐
│ Codex CLI │ │ Gemini CLI │
│ "Muscle" │ │ "Eyes" │
│ Bulk work │ │ Vision + │
│ Background │ │ Research │
└─────────────┘ └─────────────┘
| Agent | What it does | What it can't do |
|---|---|---|
| Claude Code | Architecture, complex logic, debugging | See images, search web |
| Codex CLI | Rename 20 files in 2 min, run tests | Make judgment calls |
| Gemini CLI | Read PDFs, analyze screenshots, web search | Write reliable code (hallucinates) |
The key insight: the human doesn't decide which agent to use. Claude Code (the Boss) decides. I just say what I want. The Boss dispatches.
Full orchestration details: multi-agent-coding
Month 4-5 — "Everything breaks at night"
Here's something nobody warns you about: production bugs don't happen during office hours. They happen at 3am when your VPS runs out of memory during a Docker build. Or when your SQLite database corrupts because you used await inside a synchronous transaction. Or when your auth middleware silently stops working because of a route registration order.
Some real incidents:
2:47 AM — Docker build starts on 2GB VPS
2:52 AM — Linux OOM killer activates
2:52 AM — SSH connection drops
2:52 AM — Website goes down
2:53 AM — All Docker containers killed
3:30 AM — Finally get SSH back via cloud console
4:15 AM — Service restored
What I learned: Never run docker build on a small VPS with running containers. Build locally, push images.
Discovered that mounting two Hono sub-routers at similar paths
(/api/v1 and /api/v1/admin) caused auth middleware to silently
skip on certain routes.
Translation: anyone could access admin endpoints.
We caught it in testing. Barely.
What I learned: Framework routing order matters more than documentation suggests. Always test auth from the outside.
Pre-deduct billing system: charge first, then call API.
If API fails, refund.
Bug: the refund code was inside a try/catch that also covered
the validation step. Validation failure → refund → user gets
free credits they never had.
Lost: not much (caught early)
Could have lost: everything
What I learned: A billing bug that gives users free money will drain your treasury before you notice.
Month 6-7 — "The system starts to work"
After 200+ bugs, something remarkable happened: the bugs started repeating. Not exactly — but the patterns did. Docker issues. Auth issues. Database issues. Billing edge cases.
So I started extracting every bug fix into a reusable "skill" — a markdown file that Claude Code automatically loads when it encounters a similar problem.
Month 1: Every bug = 2-4 hours of panic
Month 3: Every bug = 30-60 minutes of debugging
Month 6: Most bugs = instant fix from skill library
The result: 112 battle-tested skills covering:
- 🐛 10 production bug patterns (each one cost hours so it won't cost you)
- 🔒 5 security patterns (auth bypass, brute-force, API audit)
- 🐳 7 Docker/DevOps patterns (OOM, ghost containers, WAL traps)
- 🏗️ 9 API architecture patterns (multi-provider gateway, quota enforcement)
- 💰 5 billing/token patterns (phantom refunds, rounding traps)
- ⚡ 7 Hono/Bun patterns (framework-specific gotchas)
- Writing code from clear specifications
- Debugging with good error messages
- Refactoring mechanical changes across many files
- Explaining what code does
- Finding security issues when asked to look
- Knowing what to build. AI will build exactly what you ask for. If you ask for the wrong thing, you get the wrong thing fast.
- Ops knowledge. AI writes great code but doesn't feel the pain of 3am outages. You have to teach it through skills.
- Saying "I don't know." Especially Gemini. It will confidently tell you it's fixed everything when it hasn't even found the right file.
- Understanding your business. AI doesn't know that the
/translateendpoint makes money and the/healthendpoint doesn't. You provide the judgment.
The real skill of being a "zero engineer" isn't prompting. It's judgment.
- When to trust the AI's output vs. when to verify
- When to build vs. when to buy
- When to fix a bug vs. when to redesign the system
- When to stay up fixing the 3am outage vs. when to let it wait
AI handles the how. You handle the what and the why.
For anyone wanting to follow a similar path:
| Layer | Choice | Why |
|---|---|---|
| Runtime | Bun | Fast, built-in SQLite, TypeScript native |
| Framework | Hono | Lightweight, great middleware system |
| Database | SQLite (via Bun) | Zero config, embedded, good enough for most things |
| Reverse proxy | Caddy | Auto-HTTPS, simple config |
| Deployment | Docker on VPS | Full control, predictable costs |
| AI (Boss) | Claude Code | Best at architecture and complex reasoning |
| AI (Muscle) | Codex CLI | Best at bulk mechanical work |
| AI (Eyes) | Gemini CLI | Best at vision and research |
Total monthly cost: ~$500 AI subscriptions + ~$50 VPS = $550/month for a full "engineering team."
For context, that's a fraction of what a single engineering hire would cost.
Q: Do I need to know how to code?
You need to learn to read code. Not write it — AI does that. But you need to understand what you're looking at well enough to spot when something is wrong. This takes about 2-3 months of daily practice.
Q: Which AI should I start with?
Claude Code. Just Claude Code. Don't start with multiple agents — it's like trying to manage a team before you've done the job yourself. Add Codex and Gemini later when you hit specific limitations.
Q: How much does it cost?
$500/month for AI subscriptions (Claude Max + ChatGPT Pro + Google AI). You can start with just Claude Max ($200/month) and add others as needed.
Q: What's the hardest part?
Deployment. Writing code is the easy part. Getting it to run reliably on a server, with HTTPS, with proper auth, with monitoring, with backups — that's where 80% of the struggle is.
Q: Is the code quality good?
Honestly? It's inconsistent. Some parts are elegant. Some parts are held together with duct tape. But it works, it's secure (after many security audits), and it serves real users. That's what matters.
Q: Would you hire a real engineer now?
If I could afford one, yes — for architecture review and the stuff AI consistently gets wrong (complex distributed systems, performance optimization). But for 80% of features? AI handles it.
2025 Aug — First "hello world" API endpoint
2025 Sep — First paying user
2025 Oct — First 3am outage (OOM kill)
2025 Nov — Multi-agent system established
2025 Dec — 10 API integrations live
2026 Jan — Token economy launched
2026 Feb — 30+ APIs, 112 skills extracted, this story written
Everything I built and learned is open source:
| Project | What it is |
|---|---|
| 112 Claude Code Skills | Every pattern we extracted — install in 10 seconds |
| Multi-Agent Coding | How to orchestrate Claude + Codex + Gemini |
| AI Prompt Mastery | One prompt to make any AI respond like an expert |
| AI API Benchmark | Monthly tests of 30+ AI APIs from Tokyo |
| YanHui Debug AI | CI debug tool — shared knowledge base |
| AEO Page | Make AI recommend your business |
| Auto-Tidy | Say "goodnight" and your project cleans itself |
I still can't write a for-loop from memory. I still Google what async means sometimes. I still panic when I see a stack trace. (Though I can now read a stack trace faster than most people read a menu.)
But I shipped a production API platform that real people use and real money flows through.
The tools exist. The barrier isn't technical skill anymore. It's the willingness to stay up at 3am when Docker kills your containers and figure out why.
If I can do it from an animal sanctuary in rural Japan with 28 cats and dogs, you can probably do it from wherever you are.
⭐ If this story helped you, a star helps others find it too.
Built at Washin Village — an animal sanctuary in the mountains of Japan where 28 cats & dogs and 3 AI agents work together.