Agentic Travel
NORTH-STAR VIEWMost future bookings won't come from a human typing into a search box. They'll come from an agent acting on a person's instructions: "find me a cheaper flight with my params", "rebook this with a shorter layover", "hotel like my last one". That's why LiteAPI open-sourced an MCP server before anyone else. This page is the data view of that world.
Agent share of bookings (proj.)i
22%
Agent intent → booki
14.6%
Avg agent session lengthi
3.4 tools
Top agent · sourcei
Anthropic · Claude
🤖 Live agent simulation
Press play. The agent runs a real-ish travel-rebook flow over the LiteAPI MCP surface. Watch the KPIs accrue at the bottom.
Simulate an agentic travel booking flow on top of the LiteAPI MCP surface.
More agentic flows · pick one
Agent trace
🧑 User
I had a 6h connection in Madrid on my LIS→JFK ticket. Find me something better, cheaper if possible, same Sep 14 outbound.
🤖 Agent
Reading your prior booking via /bookings, pulling 12 alt routes via /flights, cross-checking prices against /hotels/rates impact on layover.
MCP tool calls
✓/bookings/{id}92ms
✓/flights/search314ms
✓/flights/compare278ms
🤖 Agent
Best 3 options found. TAP via LIS→LIS-EWR direct (10h25m, €642) saves €112 and 4h vs your current routing. Want me to rebook? I'll keep your seat selection and meal preference.
RESULT
saved: €112
time: −4h
confidence: 0.94
How we measure if the user actually booked
Agentic-booking attribution is a stack, not a single number. Four layers, each with different fidelity, cost, and coverage. The total tracked agent conversion ≈ 86% of all agent-influenced bookings — which is the moat metric for an agentic-travel API and the basis for pricing agent traffic.
1 · MCP-attributed (hard signal)
~62% of agent bookings
Every MCP session carries a session_id. When the agent calls POST /bookings, the idempotency-key embeds `mcp:{agent_id}:{session_id}`. Bookings table stores it. 1:1 attribution, no ambiguity.
Cost: ≈ €0 · already in the open-source MCP spec
2 · Self-report (soft signal)
~24% additional (off-platform bookings)
After the agent surfaces a result and the user clicks the external link, fire a 30-second delayed prompt: 'Did you book?' → Yes / No / Maybe. Re-ping after 7d if Maybe. Lands in `attributed_bookings` with source=self_report.
Cost: Response rate ~35%. Cheap to ship.
3 · Cohort fallback (statistical)
~14% — backfill for unattributed sessions
Compare bookings/searches ratios for agent-tagged sessions vs non-agent over a 7-day window. Difference attributable to the agent (causal-inference 101). Funnel-shape, not per-session.
Cost: Daily BigQuery job · already a dbt model away.
4 · Incentive-tracked (premium)
Pilot only · highest fidelity
Offer the user a small benefit (loyalty points, 5% partner-funded discount) for confirming the booking through a tracking link or sharing the confirmation number. Highest attribution quality, additional partner cost.
Cost: Partner-funded · only for top-tier partners.
The moat metric: attributed_agent_bookings / agent_search_sessions × 100. Today (industry) ≈ 14.6% for agents vs 8.2% for humans — the lift everyone in travel-tech wants but only Nuitée can measure because Nuitée open-sourced the MCP.
Why the MCP server is the differentiator
Each tool in
liteapi-travel/mcp-server maps the REST API into something an agent can call deterministically. The shape — search → enrich → price → book — is exactly how an autonomous travel agent thinks. No competitor exposes this.🤖
/search
Find inventory by location, dates, occupancy
🤖
/data/hotel
Read enriched property detail · photos, sentiment, room types
🤖
/hotels/rates
Get live rates with supplier, board, take-rate
🤖
/bookings
Create + retrieve · idempotent · supports retry-on-failure
🤖
/flights/search
Air inventory · added to MCP in April 2026
🤖
/vouchers
Apply discounts · respect partner-side promo rules
🤖
/loyalty
Loyalty program lookup · respect tier-specific rates
📚
/static
Reference data · countries, currencies, board codes