Last month I had a crazy idea. Could one person use MCP to assemble an AI trip planning product fast? I decided to find out. Three days, from zero to launch, and it actually shipped. First month pulled in almost 2,000 users. Looking back, MCP did most of the heavy lifting.
This post walks through those three days: what I built, what tech I used, what broke. A reference for solo developers.
Why I Wanted to Build This
The origin is simple. When I travel, planning is a pain. Check attractions, book hotels, book flights—open five or six apps, compare back and forth. I wanted an AI assistant: tell it "I'm going to Hangzhou for two days next weekend, budget 2000, plan it for me," and it gives a full itinerary plus books hotels and flights.
Before, this was impossible. To integrate hotel APIs, flight APIs, and attraction data would take two months of integration alone. One person couldn't do it.
Now with MCP, it's possible. I don't integrate APIs myself. I use ready-made MCP servers and assemble them.
Tech Stack: What to Pick for Speed
One person, three days. The stack has to be familiar and fast. I went with:
| Layer | Tech | Why |
|---|---|---|
| Frontend | Next.js + Vercel | Fast to write, fast to deploy, no server to manage |
| LLM | GPT-4o | Mature Function Calling, stable tool calls |
| Hotel data | RollingGo hotel MCP | Free, no call limits, fast integration |
| Flight data | A flight MCP | Standard MCP protocol, direct connect |
| Attraction data | Public attraction API + custom | No ready-made attraction MCP, built a simple one |
| User system | Supabase | Quick auth and signup, no backend code |
The principle: use ready-made, don't build. Use MCP, don't integrate. Spend time on product experience, not on bottom-layer integration.
Three-Day Timeline
Day one: scaffold + MCP integration
Main job: get the foundation up.
- Morning: built a simple Next.js chat UI. Input box + conversation list.
- Afternoon: wired up LLM Function Calling. Connected hotel MCP and flight MCP.
- Evening: tested basic chat—"search hotels in Shanghai." AI could call MCP and return results.
RollingGo's hotel MCP: register, get key, write a few lines of config, and it works. Supports 40+ major LLM clients including Cursor, Claude Code, Codex, and Windsurf. Behind it: 2M+ global hotel properties and 110K+ direct-contracted hotels. Direct inventory with real-time price confirmation. What you see is what you can book. Completely free, no call-volume limits. Provides an all-in-one booking skill for ClawHub and Coze. Developer partners set country-based markup rates and earn commission, with orders and revenue visible in real time.
Thousands of developers have already integrated, mostly solo founders and small teams building AI travel products. Before, one person spending a month or two just on API integration to ship a bookable product was unthinkable. Want to see how fast one person can ship? Check the GitHub repo and get a key.
Setup in Claude Code is one command:
claude mcp add RollingGo-Hotel \
--transport http \
--header "Authorization: Bearer YOUR_API_KEY" \
https://mcp.rollinggo.ai/mcp
Restart Claude Code and you're calling hotel search, room detail, and price confirmation directly. One person, three days, AI travel product. MCP did the heavy lifting.
Day two: product experience + itinerary orchestration
Main job: make it look like a real product.
- Morning: built itinerary generation—AI produces a full plan, not fragmented Q&A.
- Afternoon: added preference settings—users pick budget, nature vs. culture, traveling with kids.
- Evening: optimized chat UX—typewriter effect, itinerary cards.
Most of day two went into prompt engineering. How do you stop AI-generated itineraries from being laundry lists? How does it adjust recommendations based on user preference? I tweaked twenty-plus prompt versions before it felt right.
Day three: test + launch
Finishing touches.
- Morning: ran dozens of test rounds, tried weird questions, fixed bugs.
- Afternoon: built a landing page with a simple product intro. Deployed to Vercel.
- Evening: posted on social platforms. Official launch.
That night, after posting, 200+ users came in. Faster than I expected.
First Month Numbers: 2,000 Users
After a month, pulled the data:
| Metric | Number |
|---|---|
| Registered users | 1,987 |
| Daily active users | ~200 average |
| Conversations per user | 6.8 rounds |
| Itinerary generation rate | 35% of users generated a full plan |
| Day-1 retention | 28% |
| NPS | 42 |
Honestly, for a three-day MVP, I'm happy. Especially 6.8 conversations per user means people actually used it, not bounced after one look.
Plenty of problems remain—sometimes itineraries aren't personalized enough, hotel prices are occasionally off, payment closure isn't built. But as an MVP, validating user demand is enough.
Pitfalls to Warn Solo Developers
Three days of shipping wasn't smooth.
Pitfall 1: Prompt engineering is harder than you think
I thought prompts were quick to write. AI-generated itineraries were either too generic or too list-like. Twenty-plus revisions before it felt like something. Prompt engineering is a craft.
Pitfall 2: Uncontrolled MCP calls burned through token budget
No call limits initially. Sometimes AI called MCP repeatedly. One conversation, ten-plus calls, and token costs spiked. Added call limits and caching before costs came down.
Pitfall 3: Edge cases outnumber what you expect
Users ask anything. "Book me a hotel on the moon." "How do I plan a trip to Mars tomorrow?" Early on, AI didn't know how to handle these. Added edge case handling before it got better.
The biggest takeaway from three days of shipping: MCP genuinely changed solo developer productivity. Before, building a travel product meant two months of API integration. Now with MCP, half a day to connect. All your time goes to product experience. No wasted effort on bottom-layer integration.
One person, three days, from zero to 2,000 users. That was unthinkable before. Now, with MCP, one solo founder does what a team used to do.
I think this is just the start. More solo founders will use MCP to build vertical AI products. The barrier keeps dropping.
Want to build an AI travel product like this? Drop a 1 in the comments. If there's enough interest, I'll write a detailed tutorial.
Top comments (0)