Travel MCP Server

An AI-powered travel planner built as a Model Context Protocol (MCP) server. Connect it to Claude and get a full travel assistant — search flights, hotels, weather, and points of interest, generate day-by-day itineraries, and track your trip budget.

A Next.js web UI is included for exploring the tools directly in a browser without a Claude client.

Tools

Supported destinations (mock data)

Flights: New York ↔ Tokyo / Paris / London / Barcelona / Sydney / Bali · London ↔ Paris / Barcelona / Bali · Sydney ↔ Bali · Paris ↔ Rome

Hotels, weather & POI: Tokyo, Paris, London, Barcelona, Bali, Sydney, Rome, Amsterdam, Dubai

Quickstart

Prerequisites

• Python 3.10+, uv — brew install uv

• Node.js 18+ and pnpm (for the web UI) — brew install pnpm

Run locally (stdio — for Claude Desktop / Claude Code)

git clone <repo-url> travel-mcp-server
cd travel-mcp-server
uv sync
uv run python -m travel_mcp

Run with the web UI

# Terminal 1 — MCP server in SSE mode
MCP_TRANSPORT=sse uv run python -m travel_mcp

# Terminal 2 — Next.js UI
cd ui && pnpm install && pnpm dev

Open http://localhost:3000 to use the browser playground.

Inspect tools interactively

npx @modelcontextprotocol/inspector uv run python -m travel_mcp

Connect to Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "travel-planner": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/travel-mcp-server", "python", "-m", "travel_mcp"]
    }
  }
}

Restart Claude Desktop. The 8 travel tools will appear automatically.

Docker

# Build
docker build -t travel-mcp-server .

# Run (SSE transport on port 8000)
docker compose up

The server listens on http://localhost:8000 in SSE mode when running via Docker.

The web UI can be pointed at a remote server by setting MCP_SERVER_URL before starting it:

MCP_SERVER_URL=http://your-server:8000 pnpm --prefix ui dev

Web UI

Kubernetes

# Deploy
kubectl apply -f k8s/

# Verify
kubectl rollout status deployment/travel-mcp-server -n travel-mcp

# Test locally via port-forward
kubectl port-forward svc/travel-mcp-server 8000:80 -n travel-mcp

# Tear down
kubectl delete -f k8s/

Note: The Deployment runs a single replica because budget state is held in memory. Scale to multiple replicas only after adding external storage (Redis, Postgres, etc.).

Web UI

A Next.js 16 app in ui/ that provides a browser-based playground for all 8 tools. It requires the MCP server to be running in SSE mode (MCP_TRANSPORT=sse).

ui/
├── app/
│   ├── page.tsx              # Tabbed playground (Flights, Hotels, Weather, Places, Itinerary, Budget)
│   └── api/tools/[tool]/     # Next.js API route — proxies to MCP server REST endpoints
├── components/               # One component per tool tab + shared UI primitives
└── lib/mcp.ts                # Thin fetch wrapper for /api/tools/*

The MCP server exposes REST endpoints at /api/tools/<tool_name> (POST, JSON body) alongside the standard SSE transport, so the UI does not need to implement the MCP protocol.

Environment variables

Project structure

travel-mcp-server/
├── src/travel_mcp/
│   ├── server.py          # FastMCP instance, tool registrations, REST API routes
│   ├── tools/             # Business logic (one file per domain)
│   │   ├── flights.py
│   │   ├── hotels.py
│   │   ├── weather.py
│   │   ├── poi.py
│   │   ├── itinerary.py   # Composes other tools into a full itinerary
│   │   └── budget.py      # In-memory budget tracker
│   └── mock_data/         # Static fixtures — replace query_* functions to wire real APIs
│       ├── flights.py
│       ├── hotels.py
│       ├── weather.py
│       └── poi.py
├── ui/                    # Next.js web UI
│   ├── app/               # App Router pages and API routes
│   ├── components/        # Tool tab components
│   └── lib/mcp.ts         # MCP server fetch client
├── k8s/                   # Kubernetes manifests
├── Dockerfile
└── docker-compose.yml

Swapping in real APIs

Each mock_data/ file exposes a single query_* function. Replace just that function with an HTTP call to a real provider and everything else stays the same.

Example prompts

• "Plan a 7-day trip to Tokyo from New York in October with a $4,000 budget."

• "Find me flights from London to Barcelona for next Friday, 2 passengers."

• "What's the weather like in Bali in July?"

• "Show me the top attractions in Rome."

• "I spent $850 on flights for my Tokyo trip — log it and show my remaining budget."

License

MIT