When we started talking with Kiwi.com, they were already thinking ahead. They saw the rise of AI assistants in travel search and knew it was only a matter of time before users would expect agents to handle the full booking journey.

The question was: how do we prepare for that shift before it happens?

Kiwi recognized the potential of MCP, and together we brought a first version to life, offering a glimpse of what booking travel through an AI assistant could look like.

Why MCP?

AI assistants have a notoriously hard time retrieving up-to-date flight information. They often rely on search engines like Brave or Bing for indexing, but flight prices are dynamic and links quickly become outdated or broken.

You might think: why not just have the LLM go directly to Kiwi.com? Unfortunately, the UI built for humans is a minefield. A nonstandard date picker, cookie banner, or hidden JavaScript make navigating slow, error-prone and inefficient (not to mention extremely expensive!).

Giving the LLM direct access to an API seems like a smarter option, and in many ways it is.

However, APIs were also not designed for LLMs. APIs can have hundreds of endpoints and return vast amounts of data, most of which is irrelevant for the task at hand. This consumes the context window without adding any value for an LLM trying to book a flight.

That’s where MCP comes in. MCP bridges the gap by presenting context (tools, resources, prompts) to the LLM in a simple, standardized format it can understand. It allows those with the knowledge of the business to choose which functionalities to expose and guide the model through the task.

Kiwi.com’s MCP server

Kiwi.com's MCP server lets users of AI chatbots (ChatGPT Pro, Claude, and Cursor) search flights and get a direct booking link through natural language interaction.

It includes a search-flight tool allowing users to get instant access to a curated list of the best flights for their trip based on: 

1. Round-trip or one-way flight

2. Origin / destination (city or airport)

3. Travel dates

4. Flexibility up to +/- 3 days

5. Number and type of passengers (adult, child, infant)

6. Cabin class (economy, business, etc.)

To save the user time, the MCP generates a booking link each time a flight is proposed, even if the user hasn’t explicitly requested it.

How it works

Once having installed the MCP server (instructions here), users can simply chat with their AI assistant, for instance, “Could you help me book a flight from Paris to Rome tomorrow?”

Based on these inputs, the LLM returns a curated set of flights based on the cheapest, shortest and what the LLM considers the best flights. The user can then modify the query by choosing a preferred airport, adding some flexibility in terms of dates, adding passengers, etc.

Today the MCP server only covers one-way and round-trip flights. Choosing a multi-destination itinerary, adding checked bags, setting a maximum duration, logging into one’s account, and other options currently available on Kiwi.com are not yet available.

What Alpic optimized

Easy one-click deployment & hosting

On the Alpic platform, we designed the deployment flow to be fast, repeatable, and developer-friendly.

The Kiwi.com server is deployed directly from a Git repository using standard build commands. Once live, it’s immediately available at a secure HTTPS endpoint and mapped to a custom domain for production use.

Behind the scenes, the platform handles TLS termination, request parsing, and transport abstraction. Tool execution runs in isolated, stateless environments with built-in DDoS protection and rate limiting. This setup allowed Kiwi.com to get their first version running quickly, iterate safely, and expose their interface to real users with minimal operational overhead.

MCP server design optimizations

A shorter path to booking

LLM context windows are a precious resource, and long URLs can quickly take their toll. To avoid gobbling up context and providing potentially broken or truncated links, we suggested creating a shortened, dedicated booking link for the flights shown by the LLM.

A little structure

Instead of returning results in free-text chat form, the MCP server instructs the LLM to show results in a table format, for easier comparison and more familiar user experience.

In search, less is sometimes better

The Kiwi.com API is capable of returning potentially thousands of flights for an LLM to process. However, the server returns only a few dozen of the “best” options according to Kiwi’s existing business logic. 

Why? Kiwi.com has spent years fine-tuning its business logic to surface the best possible options. The LLM doesn’t need to reinvent the wheel by analyzing thousands of flights on every route when Kiwi has already done the hard work.

In addition, models don’t yet consistently understand price–quality tradeoffs. A well-meaning LLM might suggest a 15-hour layover in Istanbul just to save €30 on a flight from Paris to Rome. It’s best to leave that kind of judgment to the pros.

Where to next?

Logs, monitoring, and usage insights

Now that the MCP server is live, Kiwi.com will need visibility into both system behavior and user interaction to be able to improve their MCP server over time. 

For that, Alpic offers logs and monitoring for detailed traces of tool calls, as well as health metrics (uptime, latency, error rates) to monitor server performance. 

More importantly, Alpic’s MCP-specific analytics offer a view of how users are engaging with the tools, including which LLM is being used and where, which tools are called most often, which tool calls are producing errors, etc. This will help the Kiwi.com teams understand what’s working, where users struggle, and how to improve usability.

New MCP server functionalities

In addition to adding the core features of the Kiwi.com website, future iterations of the MCP server could incorporate other elements of the protocol to further refine the experience:

• Proactively asking for user input: Thanks to a June 2025 feature called “elicitations,” an MCP server can proactively ask for clarifying information to more quickly arrive at the best result for the user.
  
  For instance, if someone is searching for a flight between two cities that each have multiple airports, the MCP server could instruct the AI assistant to ask the user if they have a preference between different airport options.
  
  

• User Preferences: One of the core promises of LLMs is hyper-personalized experiences. An LLM should be able to remember and store user preferences: a window seat, a specific airline, or favoring comfort over cost. Providing tools in the MCP server to facilitate this storing and retrieving of preferences is another potential extension.

Client-side capabilities

Lastly, improvements on the MCP client side will be key to unlocking real agentic experiences for consumers. Removing installation friction via MCP server registries or “app stores” (for instance, https://claude.ai/directory) will help users adopt these technologies more easily.

In addition, instead of redirects to external links, we may soon see AI assistants capable of managing the entire booking process.

To enable this kind of automation, capabilities would need to be built out on both the client and server sides. The MCP server, for example, would need to handle logging into the user’s Kiwi.com account, accessing payment information, and proposing adjacent services, all in a secure and user-consented way.

Final thoughts

Kiwi.com knows that AI-native access is going to reshape how travelers interact with flight search and booking platforms. By investing early in an MCP server, Kiwi is helping define how access to their service could look. As travelers’ habits shift, they will be well-positioned to build out the capabilities of their MCP server and continue to meet consumers where they are headed.

If you're exploring MCP and are looking for a one-stop shop to get into production quickly, we’d love to talk!