Kiwi.com flight search MCP server
mcp-install-instructions.alpic.cloud
mcp-install-instructions.alpic.cloud
I can spend my time checking over a few days to see what the game is. The LLM Agent is gonna need a cron job system....
Otoh, LLMs booking flights? Airlines are going to exploit the ever loving shit out of that to jack rates for people not paying attention by detecting LLM queries which is completely legal.
It’s still ok for finding some routes or ideas, but far from best for cheapest (ie it cant find cheapest country to book from like skyscanner).
And don’t even get me started on business flight search - that’s basically conspiracy from airlines for you never to find deals.
"find a flight for me, wife, 2 kids aged blah blah, to <destination>, i have family in <layover cities> so at most 1 layover, always avoid United airlines or anything operated by them, because I'd rather sail on a wooden plank across the atlantic like Greta Thunberg than fly United again."
into a bunch of filters with 40 clicks.
And book! That's a very exciting bit of tooling to add into an assistant, but there's lots of complexities...
Each result includes a booking link directly to the flight chosen
Oh well, we're not at the future we hoped for yet then, but it's progress.
- With MCP Elicitations (https://modelcontextprotocol.io/specification/draft/client/e...) the server will be able to secure ask for complementary info like your name, passport and even your payments details.
- With mcp-ui (https://github.com/idosal/mcp-ui), supporting MCP Clients could allow the server to re-create the same booking flow UI that you find today on an aggregator's website
> Servers MUST NOT use elicitation to request sensitive information
For any LLM based flows to gain enough trust by the general public to handle flows that involve money, esp large sums of money, we need the equivalent of “pad lock on my browser means I’m secure” level of something easy to understand and teach everyone to see.
I suspect the argument is going to get tricky, but the business interest will win out.
Using Claude Code or your IDE of choice to book a hotel is a fun unintended side effect of this.
- Their AI can't figure out your departing city from the prompt, you have to enter it manually. - Shows you only ONE deal and no other options. - Dumps you into a pre-filtered Google Flights link that adds literally no value beyond normal search.
It will take an awful lot for me to trust an AI with an airline ticket.
Given this, I'm not sure what business purpose there is to ship an MCP API like this, aside from goodwill and exposure.
I see the same thing happening with these MCPs. Currently they’re built to ride the AI bandwagon, but when people start using them for something actually useful and engagement starts going down they’ll cripple its capabilities to restore the balance.
That ship has sailed and now consumer expectations are pretty set on “free” for a lot of things.
However this is probably still just a minority, and most people are less annoyed by ads than by having to pay however little. I'm not sure that the market correctly prices their eyeballs though, relative to those who are willing to pay micropayments. Intuitively, I'd guess the second type are probably more willing to buy things online in general. But probably the numbers and engagement metrics are prioritized more as KPIs for promotion etc.
This is why YouTube search also went to trash. If the search result list doesn't have the thing you're looking for, you might close the window. But if they intersperse some clickbait in the list, you may click that instead and stay on the platform.
The best in this is of course Tiktok, where the overwhelming usecase isn't even searching, just the for you page and tuning in to the linear stream they serve up. If the user has time to think and feels in control, they may use that control to quit the app.
It's a lot easier to inject ads into MCP than APIs.
Or much more likely OpenAI/anthropic/google will be the gatekeepers of what advertising is injected into the user’s chat.
And so on and so on.
For example, everyone thought solving context and RAG meant Vector Databases. It’s analogues to things we used to understand (hey, we need a databases, duh). Forget what you know, and be ready to throw it all out.
It’s silly to think we’ve agreed on anything other than the OpenAI API format, which even that, is just a simple HTTP call, with a simple expected formatted response.
Premature would be the word.
If the MCP server supports booking flights then they can make money from this
That said, this only makes sense if the data provided by the API is proprietary in some way. If its free or open source data, being the mcp provider likely won't provide much business value aside from the insight as to what the LLMs/users are searching for.
Now none of this is meant to excuse the behaviour of all these large platforms for all the terrible practices they engage in. But at the same time, we never figured out how to safely deal with the power exposed by these APIs.
feedback-to-devs sends feedback to the Kiwi mcp server devs.
That's an interesting way to collect feedback, but I also wonder if users of this will miss seeing that tool enabled and then inadvertently send other feedback or private data to the Kiwi devs too.
The intersection of "informed enough to use an MCP server" and "unwise enough to purchase from Kiwi" is small.
It’s equivalent to the in-browser Artifact workflow but that’s kinda annoying to work with, I usually want to export those outside of the chat client at some point.
Having an MCP server allows them to jump on the AI bandwagon while being pretty useless and not a threat to any “engagement” some of their employees are measured on.
Anyone who books their flights entirely through a chatbot deserves to have their plans screwed up, and whatever company sold the chatbot to them deserves to have these irate customers yelling at them.