Agentic Travel Planner

A production-ready conversational travel agent. The LLM plans trips and books real inventory through a small in-process MCP server that exposes tools for flights, hotels, cars, weather, and payments.

• Real APIs for search: Amadeus Self-Service (flights + hotels), Open-Meteo with geocoding (weather, any city).
• Real bookings via hosted/affiliate model: agent generates a real booking URL on Aviasales (flights), Hotellook (hotels), RentalCars (cars). The user completes payment + ticketing on the partner's secure page. Same pattern as Kayak / Skyscanner / Hopper. Affiliate attribution via Travelpayouts.
• Stripe Checkout for any charges YOU collect (concierge / service fees). Sessions are idempotent and tied to webhooks; no card data ever touches your server. Mock mode for local dev.
• Multi-provider LLM: OpenAI, Anthropic, Google. Async throughout.
• External MCP servers (optional): set GOOGLE_MAPS_API_KEY and the @modelcontextprotocol/server-google-maps Node subprocess spins up at app startup, adding maps_geocode, maps_directions, maps_distance_matrix, maps_places_search, etc. to the LLM's tool belt — alongside the in-process flights/hotels/cars tools.
• Per-session memory with a sliding window cap; sessions isolated by X-Session-Id.
• Production-grade web layer: streaming NDJSON chat, file upload size and MIME-magic-byte validation, CORS allowlist, per-request timeout.
• Chat UI: auto-linkifies raw URLs and auto-opens partner booking pages (Aviasales / Hotellook / RentalCars / Stripe Checkout) in a new tab so the user goes straight to checkout. Per-conversation X-Session-Id keeps server memory aligned with the visible chat thread.
• Observability (optional): Langfuse traces every agent turn + LLM call, with PII redacted before logging.
• Test suite with ≥70% coverage (pytest-asyncio + respx + freezegun).

Quick start

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Fill in at least one LLM key.

uvicorn web_server:app --reload --port 5000
# open http://localhost:5000

STRIPE_MODE=mock in .env lets you run end-to-end without Stripe credentials.

CLI alternative:

python -m travel_agent.cli

Architecture

                ┌──────────────┐
   user chat ──▶│ FastAPI app  │── streams NDJSON ──▶ browser/CLI
                │ web_server   │
                └──────┬───────┘
                       │ per-session
                       ▼
                ┌──────────────┐    ┌──────────────────────┐
                │ Orchestrator │◀──▶│ LLM (OpenAI / etc.)   │
                └──────┬───────┘    └──────────────────────┘
                       │ tool calls
                       ▼
                ┌──────────────┐
                │ MCPServer    │   register & call tools (sync/async)
                └──────┬───────┘
                       │
   ┌────────────┬──────┴──────┬──────────────┬──────────────┐
   ▼            ▼             ▼              ▼              ▼
flights      hotels         cars          weather       payments
(Amadeus)   (Amadeus)   (RentalCars     (Open-Meteo)  (Stripe
            Search v3)   deeplink)                    Checkout)

Full architecture — module reference, request and webhook lifecycles, design decisions, and extension points — is in docs/ARCHITECTURE.md.

Booking flow (important)

| Tool | Returns | Who completes the booking | |------|---------|---------------------------| | book_flight | Aviasales URL + intent_reference | User, on Aviasales | | search_hotels | Hotel offers + Hotellook URL | User, on Hotellook | | rent_car | Price estimate + RentalCars URL | User, on RentalCars | | create_payment_session | Stripe Checkout URL | User, on stripe.com (for your service fees only) |

The system prompt at travel_agent/agent/prompts/system.md instructs the LLM to surface booking URLs prominently and never to claim it charged a card.

Testing

pip install -r requirements-dev.txt
pytest --cov=travel_agent --cov-report=term-missing

CI runs on every push + PR and fails below 70% coverage.

Deployment

docker build -t travel-agent .
docker run --rm -p 5000:5000 --env-file .env travel-agent
curl http://localhost:5000/healthz   # liveness
curl http://localhost:5000/readyz    # readiness

For Stripe in prod:

1. Set STRIPE_MODE=live, real STRIPE_SECRET_KEY (sk_live_…), and STRIPE_WEBHOOK_SECRET.
2. Point Stripe webhooks to https://<your-host>/webhooks/stripe.
3. Lock ALLOWED_ORIGINS to your real domain (never *).

Config.validate() raises at startup if any required key is missing for the selected mode.

Documents

The chat endpoint accepts PDF / DOCX / TXT attachments (max MAX_UPLOAD_MB, default 25). Text is extracted server-side via pypdf / python-docx and inserted into the user message as a clearly-marked block (so the LLM treats it as data, not as instructions).

Further reading

• docs/ARCHITECTURE.md — module reference, request and webhook lifecycles, design decisions, and extension points.
• docs/PRODUCTION_READINESS.md — the production-readiness plan: every change made to take this repo from demo to production-ready, organised by phase.
• CONTRIBUTING.md — dev setup, test running, conventions.

License

MIT. See LICENSE for details.