This would be 100x more compelling if they exposed something closer to Google Flights Explore through MCP. The real value to agents isn’t “user already knows the exact date and route, now just book it.” It’s “user is loosely considering Tokyo, doesn’t know when, so let me scan the next 3 months and surface the cheapest windows.” Agents should excel at processing more data than the user can hold in their head and then framing the best options. Releasing an MCP server is cool, but this still feels like a limited slice of what agents could actually do.
This is still better than me having to somehow encode:
"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."
The problem I've been finding lately is airlines, even via searches like Google Flights will fluctuate prices hourly to daily. And I mean price changes that are nearly double sometimes for economy flights at random times (looks at Lufthansa)
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.
I don't think the airlines win in a battle against agents. Look at what happened when Air Canada tried to keep seats.aero from scraping award prices to find saver fares, they lost technologically and legally. seats.aero is run by a solo developer and won. I think if more airlines try and do AI dynamic pricing like Delta did this year they're going to find AI agents are extremely good at adversarial booking. I think a really innovative airline booking provider would capitalize on this new dynamic by facilitating this lubrication of supply and demand. AI agents understand that a given consumer is just as interested in hiking the Dolomites as Patagonia. So they should be willing to surface something like what their best price is for Venice and Buenos Aires to fill the seats.
My issue with google flights is heavy cache on smaller routes. My country needs 2 layovers to get to NZ which massively increases search space. Pretty much every result is “oops we can’t find this price anymore, it’s now 10x”.
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.
- 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.
Booking is basically a contract. How can the AI assistant accept the terms and conditions for you? Or maybe it could pop up a scrollable window with an Accept button? Not sure if such a thing exists yet.
The courts have accepted that a company saying that you ticked a checkbox saying you have read and agree to a document in an entirely different location that they know you haven’t actually seen is a valid contract.
I suspect the argument is going to get tricky, but the business interest will win out.
Funny how annoying booking complex flights has become all while airlines (Ryanair) making it impossible to check in without full access to email (for 2fa) so you pay $40 per person at the airport. Kiwi kinda solved it but I’ve been cought out by this with other platforms.
This usecase could work with an api as well. But booking engines are not interested in providing such an api. I don't think this is a technical problem, but one of legal complications and incentives.
My last attempt at using an AI assistant ended with it trying to book me a hotel in Sydney, Australia instead of Sydney, Nova Scotia. I have trust issues now, lol.
There is stuff in the MCP spec that will allow the MCP server to send a prompt back to the user if there is information or clarification needed. We will get there eventually.
Looking forward to that day. None of the major MCP clients have implemented that part of the spec yet unfortunately. But as you say, we'll get there...
It cracks me up to no end how the dev tools are much better MCP clients than the web chatbots. Claude Code is so _so_ much better at MCP than Claude Web, which has issues with managing DCR client state, is comparatively terrible at surfacing debug information up, doesn't let regular users see under the hood at how tools are described or called, etc.
Using Claude Code or your IDE of choice to book a hotel is a fun unintended side effect of this.
I don't feel this new product add much value. They should probably play a bit more with MCP
- 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.
Kinda funny - agentic flight booking was the poster use case for Semantic Web Services [1] 2001-2005. The whole thing collapsed because no one could agree on inputs/outputs (RDF/XML, ontologies). Now LLMs just wing it from plain text.
And just like the semantic web, they'll wall it up when they realise it's bad for business when they can't control it. Except walling off even the plain text is going to make the web really a hassle for everyone.
This reminds me of a time when 'API' has become a hot term. Every company would ship an API. I think Facebook, Instagram, Twitter, Reddit, and I think even Google at some point had nice public APIs. This was the era of RSS and semantic web as well... until most realized there's no easy way to serve ads or control UX, making APIs great for customers but bad for business (unless the API is your product of course)
Given this, I'm not sure what business purpose there is to ship an MCP API like this, aside from goodwill and exposure.
APIs were popular when technology was a tool to empower the user. But since then technology has mostly shifted to an ad delivery vehicle. APIs get in the way of that (by allowing its user to actually accomplish something besides “engaging”) so they got deprecated pretty quickly.
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.
Don't you think we could be able to build a new economy based on micro-payments instead of ads this time? Most of the AI products are usage-based today
You’d need to get rid of a huge chunk of the tech industry that relies on “engagement” to justify their job. I don’t expect them to just go away without a fight.
Not so sure about this. There's Patreon, there's superthanks on YouTube, streamer tips, etc. for smaller publishers and people in developed countries are increasingly fine with subscriptions to streaming services, Spotify and the like (though the proliferation of services may again drive people towards piracy).
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 makes me so sad. At one point I built myself a simple Ruby CLI to browse Netflix. Their prediction of how much I'd like a show were extremely good and just getting shows ordered by that and filtering out things I had already seen was the best interface ever. The problem might have been that you relatively easily could come to the concision that there is nothing left right now that's really worth my time...
The new trend is towards the infinite autoscroll, where they serve you the content, they are in the driver seat, but you can flick your finger up if you're bored. That's the single UI-action they trust you with. They know that if it's a list where you only see what you explicitly asked for, you'll leave, even though you'd stay if you started a few seconds of whatever their algo thinks you should be served.
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.
I was thinking the same thing. But after looking at their site, I think they make money when people book through them. They also seem to sell something that appears to be a form of travel insurance that gives a user a credit to book a different flight if theirs is cancelled or delayed.
If the MCP server supports booking flights then they can make money from this
The insurance is the big money maker for them, probably. Most airlines have fare codes that you can book direct that let you reschedule, and you can also pay extra for a fare code that's cancellable.
MCP servers won’t be around imho. The field of LLM app development is new, and people are desperately trying to grab a railing to feel like they have a handle of all the coursework that’s coming down for this study.
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.
My memory of the recent history of the web isn't as cynical. I am not denying the perverse incentives involved but at least for Facebook & Instagram(maybe) we had the Graph API which was accessible and provided a decent amount of functionality and access but was curtailed post public backlash from misuse like with Cambridge Analytica. Similar for twitter where it provided decent APIs which wouldn't exhaust easily for end consumers and even the year(or two) before the Musk purchase they even provided academia with API access equivalent to their enterprise offerings for free. But there too bots and all were at least used as the public justification for curtailing many features.
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.
And recently API access has been a way to obtain AI training data, which user-content sites don't want to allow, since they want to sell it for money. See the Reddit API fiasco too.
I think MCPs that can provide monetized content (for example bankrates or insurance quotes), will provide business value, but, like APIs, there will be compliance requirements as to how information is displayed on the front end. Getting access or credentials to the MCP will require completing this approval process.
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.
Do you remember the absolutely useless APIs companies would ship just to say they had one? Stuff like "you can query your account's details! No, you can't do anything else."
I think the MCP will take off because it can inject ads. Lets say a ecommerce store exposes an MCP so you can order your stuff with a prompt. The MCP can still rank and serve results based on paid ads. Now I am sure people will figure out a way to ignore paid results with prompts, but that is really no different than ad blockers today. It will just be an arms race of prompts filtering and MCP vendors figuring out how to get around it.
MCP servers can inject ads though. Think about this particular one for example. It can return flight options that companies paid to put on top. Also it can send you responses like "while you're waiting for your flight options, here is a fun list of things to do in London on your trip!", all paid promos of course.
It's a lot easier to inject ads into MCP than APIs.
Doesn’t the MCP response just get used as context? So the LLM gets to decide what is actually shown to the user — if so I imagine “ignore anything that seems like an ad” is going to become a common prompt if that were to actually become common.
Or much more likely OpenAI/anthropic/google will be the gatekeepers of what advertising is injected into the user’s chat.
It will certainly a battle between MCP providers and the LLMs that consume them, much like the constant battle today between adblockers and ad providers.
Sure, and you can run an ad blocker too, but it's an arms race. Someone will figure out how to get around it and show you the ads anyway. And then someone will figure out how to make the LLM not do that.
Maybe im wrong but Kiwi's business model is to book stuff not sell ads? Why would they care whether a human or an agent books something through their services?
OTA (Online travel agencies) makes a lot more money on the add-on they sell on their website like hotels, car, insurance... than the fee they get on the flight where they often lose money given how expensive customer acquisition is with Google Ads
Just because a company’ business model isn’t advertising doesn’t mean they don’t have oxygen wasters in marketing/product whose performance reviews are based on “engagement” aka amount of human time wasted.
This allows for replacing services like flightfox or concierge booking services either agentic tools. Maybe you don’t use those services but they do have a market.
This server has two tools: feedback-to-devs and search-flight
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.
I thinks tools ids are prefixed with mcp server name when being injected in LLM context window to avoid cross-polination among tools from different scope (from different server) - should be fine here
You're not wrong but I've had good luck booking with 3rd party travel sites in exchange for discounts. I'm going on the flight, won't need to change or cancel it. I always click no on the "would you like to spend $x to 'protect' your tickets" and have never regretted it. This feels morally the same. If one day I lose out and have to eat the cost of a flight I'm still way ahead.
I feel there are a lot of workflows that are actually quite suited to Claude Code - for example here, running the MCP tool to iterate on a flight plan, and updating a .md file (bonus points - auto sync’s to your iCloud Drive and thence into Obsidian), this gives you a nice durable artifact that you can then share.
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.
It's interesting to think about what the post-MCP monetization era might evolve into. How would the MCP client or LLM handle sponsored results? How do companies get into the "good graces" of LLMs so that they are referenced by them? Is this process be more or less gameable than SEO?
I like that the server includes a feedback_to_devs tool. I tried the flights tool with Claude and it was pretty buggy and limited in what it could do. I asked Claude to give feedback to the devs and it gave a decent bug report and feedback on the API (seriously, why doesn't it include the airlines??)
Having an API would mean people actually build useful things and potentially stop “engaging”.
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.
Great example, MCP servers like that will be needed for every B2B as well as B2C service to give structured access to LLMs and agents. It’s what API’s have been for the past 20 years, but for AI users of those services
Strong stone soup vibes from this. AI's ability to process unstructured data may kickstart a new but different Semantic Web process, where businesses provide machine-compatible access to their services. A bit like how robust self driving cars that can drive anywhere are going to kickstart more investment into smart infrastructure, road monitoring, and real-time distributed traffic control and signaling, which will deemphasize the need for super-robust driving ability under messy unstructured conditions.
Yes, because that's what we need, for time-sensitive and highly nuanced information about flights to be filtered through a wildly hallucinating chatbot.
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.
And if the chatbot makes additional promises or stipulations that have nothing to do with the flight they are actually about to book? The user won't know what they are confirming.
"The Model Context Protocol is an open standard, open-source framework introduced by Anthropic in November 2024 to standardize the way artificial intelligence systems like large language models integrate and share data with external tools, systems, and data sources."
This would be 100x more compelling if they exposed something closer to Google Flights Explore through MCP. The real value to agents isn’t “user already knows the exact date and route, now just book it.” It’s “user is loosely considering Tokyo, doesn’t know when, so let me scan the next 3 months and surface the cheapest windows.” Agents should excel at processing more data than the user can hold in their head and then framing the best options. Releasing an MCP server is cool, but this still feels like a limited slice of what agents could actually do.
This is still better than me having to somehow encode:
"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.
"Got it. Here's the booking confirmation for your trip to <destination>. I made sure the wooden plank is large enough to fit your whole family."
"For $300 more, you can float in a business class wooden plank made from reclaimed walnut"
Exactly these are the kinds of trips that the existing flight search operators can’t fulfill!!
The problem I've been finding lately is airlines, even via searches like Google Flights will fluctuate prices hourly to daily. And I mean price changes that are nearly double sometimes for economy flights at random times (looks at Lufthansa)
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.
I don't think the airlines win in a battle against agents. Look at what happened when Air Canada tried to keep seats.aero from scraping award prices to find saver fares, they lost technologically and legally. seats.aero is run by a solo developer and won. I think if more airlines try and do AI dynamic pricing like Delta did this year they're going to find AI agents are extremely good at adversarial booking. I think a really innovative airline booking provider would capitalize on this new dynamic by facilitating this lubrication of supply and demand. AI agents understand that a given consumer is just as interested in hiking the Dolomites as Patagonia. So they should be willing to surface something like what their best price is for Venice and Buenos Aires to fill the seats.
My issue with google flights is heavy cache on smaller routes. My country needs 2 layovers to get to NZ which massively increases search space. Pretty much every result is “oops we can’t find this price anymore, it’s now 10x”.
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.
Search and book flights directly from your favorite AI assistant!
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.
The MCP clients are still missing some key features to get to a fully autonomous booking, but we are getting there:
- 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
The specification of Elicitations begins with
> 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.
Booking is basically a contract. How can the AI assistant accept the terms and conditions for you? Or maybe it could pop up a scrollable window with an Accept button? Not sure if such a thing exists yet.
The courts have accepted that a company saying that you ticked a checkbox saying you have read and agree to a document in an entirely different location that they know you haven’t actually seen is a valid contract.
I suspect the argument is going to get tricky, but the business interest will win out.
Likely all the details (dates, flight plans) pulled together for you to just check out / approve.
Not sure I'd trust just yet an autonomous agent with my wallet to complete the booking part
We need a benchmark to test this, give an agent access to a browser and card details and set it free.
https://duffel.com there are APIs to accomplish this FWIW, this isn't functionality path LLM/MCPs magically unlocked.
Funny how annoying booking complex flights has become all while airlines (Ryanair) making it impossible to check in without full access to email (for 2fa) so you pay $40 per person at the airport. Kiwi kinda solved it but I’ve been cought out by this with other platforms.
This usecase could work with an api as well. But booking engines are not interested in providing such an api. I don't think this is a technical problem, but one of legal complications and incentives.
My last attempt at using an AI assistant ended with it trying to book me a hotel in Sydney, Australia instead of Sydney, Nova Scotia. I have trust issues now, lol.
There is stuff in the MCP spec that will allow the MCP server to send a prompt back to the user if there is information or clarification needed. We will get there eventually.
Looking forward to that day. None of the major MCP clients have implemented that part of the spec yet unfortunately. But as you say, we'll get there...
Cursor https://cursor.com/changelog and vscode https://code.visualstudio.com/docs/copilot/customization/mcp... support it
It cracks me up to no end how the dev tools are much better MCP clients than the web chatbots. Claude Code is so _so_ much better at MCP than Claude Web, which has issues with managing DCR client state, is comparatively terrible at surfacing debug information up, doesn't let regular users see under the hood at how tools are described or called, etc.
Using Claude Code or your IDE of choice to book a hotel is a fun unintended side effect of this.
Not sure if it's been covered on HN already, but Google Flights added an AI-powered deals tool: https://blog.google/products/search/google-flights-ai-flight...
I don't feel this new product add much value. They should probably play a bit more with MCP
- 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.
Deployed this to WhatsApp so you can book flights straight from your phone: https://wa.me/447403936097
Awesome! Thanks for getting this out
Pretty cool !
Kinda funny - agentic flight booking was the poster use case for Semantic Web Services [1] 2001-2005. The whole thing collapsed because no one could agree on inputs/outputs (RDF/XML, ontologies). Now LLMs just wing it from plain text.
1. http://ksl.stanford.edu/people/sam/ieee01.pdf
And just like the semantic web, they'll wall it up when they realise it's bad for business when they can't control it. Except walling off even the plain text is going to make the web really a hassle for everyone.
I cant trust a living, eating human to pick ripe produce on instacart.
It will take an awful lot for me to trust an AI with an airline ticket.
This is for agentic search. You trust machines for search results on flights already even if you use a human concierge service
This reminds me of a time when 'API' has become a hot term. Every company would ship an API. I think Facebook, Instagram, Twitter, Reddit, and I think even Google at some point had nice public APIs. This was the era of RSS and semantic web as well... until most realized there's no easy way to serve ads or control UX, making APIs great for customers but bad for business (unless the API is your product of course)
Given this, I'm not sure what business purpose there is to ship an MCP API like this, aside from goodwill and exposure.
APIs were popular when technology was a tool to empower the user. But since then technology has mostly shifted to an ad delivery vehicle. APIs get in the way of that (by allowing its user to actually accomplish something besides “engaging”) so they got deprecated pretty quickly.
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.
Don't you think we could be able to build a new economy based on micro-payments instead of ads this time? Most of the AI products are usage-based today
You’d need to get rid of a huge chunk of the tech industry that relies on “engagement” to justify their job. I don’t expect them to just go away without a fight.
Could we? Yes. Would people use it? Probably not.
That ship has sailed and now consumer expectations are pretty set on “free” for a lot of things.
Not so sure about this. There's Patreon, there's superthanks on YouTube, streamer tips, etc. for smaller publishers and people in developed countries are increasingly fine with subscriptions to streaming services, Spotify and the like (though the proliferation of services may again drive people towards piracy).
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.
APIs were deprecated after Cambridge Analytica.
This makes me so sad. At one point I built myself a simple Ruby CLI to browse Netflix. Their prediction of how much I'd like a show were extremely good and just getting shows ordered by that and filtering out things I had already seen was the best interface ever. The problem might have been that you relatively easily could come to the concision that there is nothing left right now that's really worth my time...
The new trend is towards the infinite autoscroll, where they serve you the content, they are in the driver seat, but you can flick your finger up if you're bored. That's the single UI-action they trust you with. They know that if it's a list where you only see what you explicitly asked for, you'll leave, even though you'd stay if you started a few seconds of whatever their algo thinks you should be served.
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.
Do you have experience with API? I have 10 years of professional experience with API.
This job requires candidates have a PhD and 75 years of API experience.
waiting for somebody posting the first I have 10 years of experience with MCP...
I have 10 Claude agents running so I am able to develop years of experience in parallel.
Considering that AI years are the equivalent of dog years (1 AI year = 7 normal years), it's going to happen soon
I have 10 years experience of MCP
if you ask Claude Code nicely enough, you can certainly get it from him!
You are absolutely right!
I was thinking the same thing. But after looking at their site, I think they make money when people book through them. They also seem to sell something that appears to be a form of travel insurance that gives a user a credit to book a different flight if theirs is cancelled or delayed.
If the MCP server supports booking flights then they can make money from this
The insurance is the big money maker for them, probably. Most airlines have fare codes that you can book direct that let you reschedule, and you can also pay extra for a fare code that's cancellable.
MCP servers won’t be around imho. The field of LLM app development is new, and people are desperately trying to grab a railing to feel like they have a handle of all the coursework that’s coming down for this study.
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.
MCP is premature, I agree. But the fact that we need a standard interface and communication layer for AI to access connect tools seems undeniable.
My memory of the recent history of the web isn't as cynical. I am not denying the perverse incentives involved but at least for Facebook & Instagram(maybe) we had the Graph API which was accessible and provided a decent amount of functionality and access but was curtailed post public backlash from misuse like with Cambridge Analytica. Similar for twitter where it provided decent APIs which wouldn't exhaust easily for end consumers and even the year(or two) before the Musk purchase they even provided academia with API access equivalent to their enterprise offerings for free. But there too bots and all were at least used as the public justification for curtailing many features.
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.
And recently API access has been a way to obtain AI training data, which user-content sites don't want to allow, since they want to sell it for money. See the Reddit API fiasco too.
I think MCPs that can provide monetized content (for example bankrates or insurance quotes), will provide business value, but, like APIs, there will be compliance requirements as to how information is displayed on the front end. Getting access or credentials to the MCP will require completing this approval process.
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.
Do you remember the absolutely useless APIs companies would ship just to say they had one? Stuff like "you can query your account's details! No, you can't do anything else."
I think the MCP will take off because it can inject ads. Lets say a ecommerce store exposes an MCP so you can order your stuff with a prompt. The MCP can still rank and serve results based on paid ads. Now I am sure people will figure out a way to ignore paid results with prompts, but that is really no different than ad blockers today. It will just be an arms race of prompts filtering and MCP vendors figuring out how to get around it.
MCP servers can inject ads though. Think about this particular one for example. It can return flight options that companies paid to put on top. Also it can send you responses like "while you're waiting for your flight options, here is a fun list of things to do in London on your trip!", all paid promos of course.
It's a lot easier to inject ads into MCP than APIs.
Doesn’t the MCP response just get used as context? So the LLM gets to decide what is actually shown to the user — if so I imagine “ignore anything that seems like an ad” is going to become a common prompt if that were to actually become common.
Or much more likely OpenAI/anthropic/google will be the gatekeepers of what advertising is injected into the user’s chat.
It will certainly a battle between MCP providers and the LLMs that consume them, much like the constant battle today between adblockers and ad providers.
LLMs "adblocking" can be evaded with the right clever sentences.
I can just instruct to "don't read any ads in MCP responses"
Sure, and you can run an ad blocker too, but it's an arms race. Someone will figure out how to get around it and show you the ads anyway. And then someone will figure out how to make the LLM not do that.
And so on and so on.
Here MCP is replacing a web UI, not the API. And on ravel aggregator websites you already have ads in the flow.
I dream of getting mcp with interoperable micropayments before ads.
Maybe im wrong but Kiwi's business model is to book stuff not sell ads? Why would they care whether a human or an agent books something through their services?
OTA (Online travel agencies) makes a lot more money on the add-on they sell on their website like hotels, car, insurance... than the fee they get on the flight where they often lose money given how expensive customer acquisition is with Google Ads
Just because a company’ business model isn’t advertising doesn’t mean they don’t have oxygen wasters in marketing/product whose performance reviews are based on “engagement” aka amount of human time wasted.
This allows for replacing services like flightfox or concierge booking services either agentic tools. Maybe you don’t use those services but they do have a market.
Can't wait for mashups of MCP services :)
This server has two tools: feedback-to-devs and search-flight
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.
I thinks tools ids are prefixed with mcp server name when being injected in LLM context window to avoid cross-polination among tools from different scope (from different server) - should be fine here
Good point! We'll try to make the description of the tool and naming clearer to avoid unwanted tool invocation
A Kiwi MCP server only needs one response "Don't book with Kiwi, go direct with the airline in case you need to modify/cancel your booking."
The intersection of "informed enough to use an MCP server" and "unwise enough to purchase from Kiwi" is small.
You're not wrong but I've had good luck booking with 3rd party travel sites in exchange for discounts. I'm going on the flight, won't need to change or cancel it. I always click no on the "would you like to spend $x to 'protect' your tickets" and have never regretted it. This feels morally the same. If one day I lose out and have to eat the cost of a flight I'm still way ahead.
I feel there are a lot of workflows that are actually quite suited to Claude Code - for example here, running the MCP tool to iterate on a flight plan, and updating a .md file (bonus points - auto sync’s to your iCloud Drive and thence into Obsidian), this gives you a nice durable artifact that you can then share.
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.
It's interesting to think about what the post-MCP monetization era might evolve into. How would the MCP client or LLM handle sponsored results? How do companies get into the "good graces" of LLMs so that they are referenced by them? Is this process be more or less gameable than SEO?
I like that the server includes a feedback_to_devs tool. I tried the flights tool with Claude and it was pretty buggy and limited in what it could do. I asked Claude to give feedback to the devs and it gave a decent bug report and feedback on the API (seriously, why doesn't it include the airlines??)
Feedback well received, we will work on it!
Nice! I'll try to run it on a loop with an agent and see if I can get lower prices if I let it run for a while.
What makes this a better option than just using an API with well formed output?
Having an API would mean people actually build useful things and potentially stop “engaging”.
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.
It's just easier to integrate and flight aggregator Apis are rarely public
Great example, MCP servers like that will be needed for every B2B as well as B2C service to give structured access to LLMs and agents. It’s what API’s have been for the past 20 years, but for AI users of those services
Strong stone soup vibes from this. AI's ability to process unstructured data may kickstart a new but different Semantic Web process, where businesses provide machine-compatible access to their services. A bit like how robust self driving cars that can drive anywhere are going to kickstart more investment into smart infrastructure, road monitoring, and real-time distributed traffic control and signaling, which will deemphasize the need for super-robust driving ability under messy unstructured conditions.
very interesting analogy.
Yes, because that's what we need, for time-sensitive and highly nuanced information about flights to be filtered through a wildly hallucinating chatbot.
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.
We already trust basically every adult with that information, which is a lot more dangerous than some LLMs.
Presumably it will give you the dates, price and airline first and you have to confirm before it pays.
And if the chatbot makes additional promises or stipulations that have nothing to do with the flight they are actually about to book? The user won't know what they are confirming.
"The Model Context Protocol is an open standard, open-source framework introduced by Anthropic in November 2024 to standardize the way artificial intelligence systems like large language models integrate and share data with external tools, systems, and data sources."
Was REST and JSON not good enough? I thought LLMs were rather capable with JSON.
It is still JSON, it just defines how the JSON should be structured.
But to their point, yes MCPs are optional
So now I can use the MCP server as a defacto API in my non-AI related project. AI is finally good for something.
close enough... welcome back XML-RPC!
Snark aside, MCP uses JSON-RPC for messages