Building an agent-native flight booking experience with Kiwi.com
B2C
Online travel booking
MCP & ChatGPT app
Kiwi.com rethought flight booking for a world where AI agents are the primary interface. By building an MCP-native experience, they made live travel search usable, reliable, and agent-friendly. With Alpic, Kiwi.com shipped one of the first production-ready flight booking services for ChatGPT & Claude.
Kiwi.com has built its reputation on one simple idea: travel should be accessible to everyone. By combining global flight content with its proprietary Kiwi Code algorithm, the company uncovers route combinations and prices that traditional search engines miss. Billions of price checks run daily behind the scenes, powering more than 100 million searches across Kiwi.com’s website and partner channels.
As AI assistants started to reshape how users plan and book travel, Kiwi.com made an early strategic decision. Rather than treating AI as a distribution afterthought, the team wanted to be one of the first travel companies to design a booking experience that works natively for AI agents.
The challenge
Kiwi.com saw a clear shift coming in how travelers would discover flights. AI assistants were becoming a primary interface, but existing approaches were poorly suited to that world.
Real-time flight data changes constantly, while language models often relied on cached or indexed information. Web interfaces are optimized for humans, not agents, and introduced friction through complex navigation and dynamic elements. At the same time, traditional flight search APIs exposed numerous endpoints that quickly overwhelmed language models when the goal was a single, relevant itinerary.
Kiwi.com identified MCP as the interface that could bridge this gap, enabling AI agents to interact with live travel services reliably and efficiently, while preserving the company’s existing business logic.
The solution
In the summer of 2025, Kiwi.com set out to build its first Model Context Protocol server as a prototype for agent-native flight search and booking. The goal was not to expose the entire API surface, but to design a focused, agent-friendly interface that mirrored how travelers actually think about trips.
Using the Kiwi.com MCP server, travelers could search and book flights through natural language, from simple point-to-point trips to flexible date searches and family travel. The MCP server translated these requests into structured itineraries with clear comparisons and direct booking links.

Only a few months later, OpenAI announced ChatGPT Apps, combining MCP with native UI. As a result, readiness for ChatGPT Apps quickly became a priority, with MCP-backed services set to become directly accessible inside ChatGPT.
This shift required rethinking the experience with a UI-driven workflow, including interactive elements such as maps and itinerary views. It meant translating Kiwi.com’s complex travel logic into tools, responses, and interface components that worked seamlessly for both AI agents and end users.
What the server supports today:
One-way and round-trip flight search
Origin and destination by city or airport code
Date flexibility up to +/- 3 days
Adult, child, and infant passenger types
Cabin class selection (economy, business, etc.)
Curated results: cheapest, fastest, and best pick
Instant booking link generated with every result
Table-formatted output for easy comparison
What Alpic optimized
Shorter path to booking
Long URLs eat context windows. We introduced shortened, dedicated booking links generated with every result, keeping conversations clean and links reliable.
Structured results by default
Instead of free-text output, the MCP server instructs the LLM to return flights in a table format, making comparison intuitive and the experience familiar.
Fewer results, better results
The server surfaces a curated shortlist using Kiwi's existing business logic, not raw API output. Kiwi spent years tuning that logic. The LLM doesn't need to reinvent it.
The Results
Idea to production in a few weeks
From the first conversation to a live server deployed on a custom domain, with TLS, DDoS protection, and rate limiting built in.
3 AI platforms at launch
The MCP server is live and usable across ChatGPT Pro, Claude, and Cursor on day one, distributing Kiwi.com where travelers are already searching.
Live conversational flight search
Travelers get real-time results and a booking link in a single exchange. No UI, no redirects, no broken links. Search to booking in one step.
0 infrastructure to manage
Alpic handles hosting, monitoring, uptime, and TLS. Kiwi's team focuses on their product, not their server stack.
"A big thank you to the Alpic team for working hand in hand to deploy and manage the first MCP server for searching and booking flights in Kiwi.com."
Mario Gavira
VP of Growth and Brand at Kiwi.com
