I Let Claude Plan My Queenstown Trip
I built a small MCP bridge between Claude and Wanderlog, asked it to plan a week in Queenstown, then went. Here's what actually saved me time on the ground.
My partner and I had a week in Queenstown coming up and I hadn’t planned any of it. Flights booked, a Frankton apartment booked, a hire car booked. That was it. No itinerary, no restaurant picks, no idea what I was packing for May in the Southern Alps.
So I asked Claude to plan it. Not in a chat I’d have to copy-paste from later. I wanted the output to land directly in Wanderlog, the trip-planning app I actually use. To make that happen, I spent an afternoon writing a small MCP server that lets Claude read and write Wanderlog trips through the same sync protocol the web app uses. A bridge, nothing more.
Then we went on the trip. We got back two days ago. This is what came back, what we actually used, and why I think this is the part of working with AI that’s quietly become real.
The bridge, briefly
MCP is the protocol that lets an AI assistant call tools that live outside the model. You expose add_place, add_note, add_checklist, and the model can use them like any other capability. Wanderlog has no public API, so I had to talk to its real-time sync layer directly. The surface area I needed was small, and once it worked, I stopped thinking about it.
That’s the whole point. The bridge isn’t the post. The bridge is the part that gets out of the way.
What it produced
I told Claude the dates, that we were going as a couple, our dietary requirements, that we wanted some hiking and some scenic driving and one real splurge dinner. It came back with a six-day Wanderlog board with 26 places, two checklists, and notes attached to almost everything. Four things on it changed how the trip went.
A day-by-day itinerary that respected timing
Each day had three to five places with start and end times attached. Not “morning: town”. Actual 09:15–09:30 windows that accounted for opening hours, drive times, and the rhythm of the day. Crown Range Summit was a 15-minute photo stop, not a destination. Mirror Lakes was 11:00–11:15 because that’s how long the Milford coach pulls in for. Big Fig in Wānaka was scheduled for 12:15 because we’d be hungry after the Albert Town outlet walk and Wānaka Station Park, in that order.
The detail I didn’t expect was the routing logic in the notes. The Wānaka day had a one-line tip: take Crown Range Road, not SH6 via Cromwell, because the former is the scenic route. We probably would have gone that way anyway, but it was good to have the call made for us before we were standing at a junction with the GPS half-loaded.
What that bought us on the ground: I never sat in the car deciding what was next. Open Wanderlog, look at the next pin, drive. The pacing was right enough that we only re-shuffled one thing all week.
Restaurant picks that accommodated dietary needs
This was the one I expected to be flaky and wasn’t. I gave it our dietary requirements, and instead of pointing at the obvious tourist pins and calling it a day, it came back with a shortlist of options that worked: a mix of certified spots, kitchens it flagged honestly as compatible-but-not-certified, and one place worth booking ahead for the splurge dinner. Each pin had a note attached: what to order, whether to book, rough price. We ate well all week and never had to do the “is this safe” dance at the door.
The thing I want to call out is what was actually happening behind that. Claude wasn’t recalling a list from training data. It was running its own research, cross-referencing certification status, reading reviews, and then writing the results straight into Wanderlog as place pins with notes. That’s the unlock. The model can go and check, and now it has somewhere to put what it found, in the same shape the next step expects.
Pre-trip checklists with real specifics
Two checklists landed in the board before we flew. The pre-trip one was ten items. None of them were “remember your passport.” They were the things I would otherwise have remembered late, in the wrong order, the night before:
The second checklist was the one that surprised me: a fourteen-item kit list specifically for the Moke Lake stargazing trip. I knew nothing about stargazing. I didn’t know the Wakatipu basin had been picking up dark-sky recognition the last couple of years, or that Moke Lake sits in a natural bowl that shields it from Queenstown’s glow. I didn’t know what a red-light headlamp was for, or that night-vision adapts over twenty minutes and one phone screen at full brightness will reset it.
This is where two things stacked. Claude already knows a lot about stargazing (preserve your night vision, sit still in cold air, the Magellanic Clouds are the southern-sky thing worth driving for). That’s general knowledge that lives in the model. But it also went and researched the specific bits: that the last stretch of Moke Lake Road is unsealed, that there’s no signal once you’re at the lake, that there’s no petrol out there. And then it wrote both kinds of knowledge into the same checklist:
- Red-light headlamp (preserves night vision).
- Stellarium Mobile with red-screen night mode.
- Full tank of fuel before leaving Queenstown. No petrol near Moke Lake.
- Late-night Fergburger on the way back. Only the central window stays open past 22:00.
That last one is the kind of thing only a thoughtful human friend usually thinks to say. It came from a tool call.
Contingency, not just plans
The thing I keep coming back to is a single note attached to the stargazing day:
STARGAZING RAIN BACKUP. May 7 forecast is 70% rain, 19.5mm. Moke Lake stargazing is weather-dependent: clear skies required. CHECK FORECAST AT 18:00 BEFORE COMMITTING TO THE DRIVE.
It had pre-checked the forecast for that day, flagged the risk, told me when to make the call, and given me a backup. I did check at 18:00. The cloud broke. We drove, and we sat by Moke Lake with a thermos and saw the Magellanic Clouds.
A plan that quietly hands you a decision tree for the days you’ll be tired and the weather will be uncooperative is doing more than a plan that just lists places.
What this is like now
The headline isn’t that an AI can produce a travel itinerary. ChatGPT could do that two years ago. The change is that the output now lands in the place I was going to plan the trip in anyway, with start times, addresses, phone numbers, dietary notes, contingency advice, and links to specific apps. I don’t have to translate, transcribe, or re-enter any of it. I opened Wanderlog on the plane and the trip was already there.
This is the bit I think people underrate about MCP. The model getting smarter isn’t the point. The point is the model being able to act inside the tools you already use, on data that’s already yours, in the format the next step expects.
A year ago I would have asked an AI for a Queenstown plan, copy-pasted the output into a Google Doc, then typed half of it into Wanderlog by hand and given up halfway through. This time I asked once, packed off the checklist, drove the route it recommended, ate at the places it found, watched the sky from a lake it told me about, and came home.
Worth the afternoon spent on the bridge. I’d build it again.
The bridge is open-source. If you use Wanderlog and want Claude (or any MCP-aware client) writing into it the way mine does, the server is here.
MCP server for Wanderlog — build and edit trip itineraries through conversation