The backstory: about a year ago I took a few spontaneous trips through Europe and didn't want to spend ages researching beforehand. I asked Claude instead, and it turned out to be a really good travel advisor. For most destinations it has a ton of good tips. The only thing that sucked was the UX. Copy-pasting place names into Google Maps one by one is tedious, and it's even worse on your phone once you're already there and just want to see what's nearby. So I built myself something that works the way I wanted, and figured it might be useful to you too.

I know Claude and ChatGPT can sort of show things on a map themselves, but it's not really usable. It's buried in the chat history, it isn't interactive, and you can't see where you actually are right now.

I started building in October, had a first version in November, and have been dogfooding it hard on my own trips since. It's at the point now where I rely on it. It factors in the weather and your current location, and it'll warn you before you book a flight somewhere you shouldn't (thanks to their training cutoff LLMs will happily recommend a sightseeing trip through the Persian golf region). You can also group several chats into a project so they share context across one trip.

There's a demo with some example conversations you can click through, or you can sign up and try it for real, you get some free messages. I had to put it behind a signup because leaving a chat window open to the whole internet would kill my wallet.

The two biggest challenges I faced were geocoding reliability and LLM costs.

For geocoding, I didn't just want pins, I wanted real geometry wherever possible. A pin labeled "Le Marais" is fine, but if I'm standing nearby I'd rather see the actual area than eyeball it. OpenStreetMap is the obvious source, but you have to fight it. Ask for a neighborhood and it'll happily hand you a random bus stop, or a POI with the same name on the other side of the city. So the geocoder grew into a whole system: multiple sources of truth, custom ranking and automated conflict resolution via LLMs.

With the LLM, I started on Sonnet 4.5 because it's great, and it (then 4.6) stayed the gold standard for a long while. A lot of the work went into keeping it cheap: maximizing cache hits and cutting tool-loop iterations, since with Claude every extra round trip means a fresh cache point and more money. Then DeepSeek cooked with V4 Pro. After a lot of testing and tuning it came out roughly on par with Sonnet at about 20% of the cost, so it's the default now. Works great, and keeps credit usage down for users too.

Happy to answer anything!