I asked Claude to book my flight & hotel, It took 6 months.
retained text
I needed to travel from Warsaw to London for an event and wanted to see how far the default Claude app on mobile/web can go if I asked it to FULLY book my whole trip. The immediate suggestions from Claude was to use the integrated connectors (Expedia, Kiwi, etc.), I tried those but they were really poor. I still had to book myself, and they didn't give full results. I knew it could do more with subagents. I started building this very quickly. I was super excited and didn't know anything. I didn't expect it to take 6 months (the date of the first commit on github). I obviously didn't end up using it for that trip, but I finished the project. https://i.redd.it/o5vb1un3ohph1.gif A week in, a friend decided joined me. We became cofounders and called it **LetsFG.** There was a ton of things to make this work: **1. Getting Offers** This was the first step, and I knew that if I didn't get ALL of the results around the internet, I wouldn't use it. I wouldn't trust it. I would still go and compare it myself to other websites just to make sure im not overpaying. This would be ruin the purpose of this project - time saving and cognitive deload. I had to get a list of all the websites that are worth checking - Claude gave me a list of about 200 websites, OTAs, metasearches, direct airlines. Then I needed to figure out how to search and find flights there. And let me tell you... Getting APIS IS IMPOSSIBLE in this industry. I have no idea why they gatekeep it so much, but anyway I ended up hosting this whole thing on my backend, where I run a separate AI agent for each website. At first I used the cheapest gemini models for this, which were ass. Later this year GPT-5.6 Luna came out which was a game changer. Regardless, this remained incredibly costly. Considering that it's just a search and I make hundreds of those, this was unacceptable. I worked on optimization for 2 weeks. I figured out that websites are just network calls, and instead of reading websites, I could just do some magic, read the network calls, and still get the full results. This lowered the cost by 20x. **2. Booking Offers** When I look back at it now this is crazy. I spent 3 months on getting end-to-end booking to work. The question with booking was can I collect the money from my card, with bank-grade safety, without Claude, or my services, or the AI agents ever seeing it. I couldn't find anything. Literally for 2 months I went from one sales call to another, talking with payment providers just to learn they can't support this use case. I went through 10 companies, and the last one I thought could support me, turned out to require a $5k setup fee, monthly minimum, and a crazy processing fees. I was lost. At the beginning of August my cousin was driving through the city with his family and they stayed for a day. He - an enigneer himself - was always curious about my startups, projects, and asked if I tried one specific company. He thought that they're startup friendly and might work. Funnily enough I thought of this exact company in the beginning, but I never reached out because I thought they'd never support this. No way. I gave them a shot. It turned out they supported my use case perfectly! They accepted the project and that started my little revolution. That company is Revolut. Sometimes a single conversation can change your situation from hopeless to absolutely amazing. **Never quit.** After having searching and booking capabilities in Claude we realised we just replaced a part of executive assistant's job. And it's in everybody's pocket. What came out later is that it can not only find us flights easier - without any cognitive work on our side - but also cheaper. On round trip flights Claude started to break the trip apart into 2 one way flights. When we thought about it, it was still unclear where the savings came from. It turns out it not only broke the flights apart, but also websites. It compared both of the flights, and choose each one from a different website. For example it booked the 1st outbound flight from Skyscanner, but the returning flight from Kayak. We tested this further on long haul flights, like Warsaw -> Beijing or New York -> Berlin, which always get at least 1 stop, and it worked exactly the same. On these flights it created offers that were even $100-250 cheaper. With this, another problem arose. What if I book a flight especially like the above, and it gets delayed or cancelled? We attacked it in 3 ways. \- First we made an algorythm that shows the reliability of every flight based on the historical data, weather, and predictions. \- Then we marked each fare that is created in the above way with "Separate tickets". For me it's still the same flow, AI books it for me, no additional work on my side. But I am informed of the risk. \- The last thing we made was a 24/7 monitoring that sends updates in case of changes. If a flight is delayed or cancelled I instantly get rebooking, replacement options. When using it in Claude, and waiting for responses we thought how can we make it more fun. How can we express the spirit our LetsFG name manifests. We figured memes. And it's beautiful. https://i.redd.it/wwc61uvg9iph1.gif There are still some things it can't do. It can't automatically rebook a flight for free, for example when the flight gets cancelled. It can suggest replacements, even suggest options that are possibly for free because they're in the same alliance. However it can't change the reservation. It can only book a new one. Can't manage a booking yet. We will work on this in the future, but today it's not able to for example cancel your flight. It needs a slight markup on the offers to pay for payment processing depending on what card I use. International cards are like 2% more. I leave the judgment if this is acceptable to you, but there is nothing I can do with it, unless I get huge volumes and negotiate with Revolut. This project was incredible to do. Feels like we're pushing the agentic commerce further. We did thousands of tests to make it work as reliable as possible. We still ocasionally find edge-cases where we have to approve bookings manually, but mostly it's solved. To shorten the feedback loops we gave Claude Code access to our machines, and Fable 5 together with Opus 5 were running the whole thing. (The token count was bigger but I changed my Plan and it reseted the streak) https://preview.redd.it/208butrg8iph1.png?width=512&format=png&auto=webp&s=2e5c9128b8b32233a12767a64d1492e60cccaac2 We hosted this thing on our website. This is what you add to Claude if you want to use it: [https://letsfg.co/developers/api/mcp](https://letsfg.co/developers/api/mcp) And this is our website if you prefer prehistoric UIs: [https://letsfg.co](https://letsfg.co) We also have a Github. We mostly use it for the community, issues, feedback and docs, and Python/NPM packages: [https://github.com/LetsFG/LetsFG](https://github.com/LetsFG/LetsFG) We're super open to feedbacks, use-cases, questions and anything. Especially about the offers finding algo, business decisions, or generally booking :D https://preview.redd.it/cgpn9qr58iph1.png?width=1901&format=png&auto=webp&s=a5ab996a593fa0c09f1271fe779063ee55b13109
document fetched 2026-09-15 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wg6ubl/i_asked_claude_to_book_my_flight_hotel_it_took_6