Thailand Fun Experiences — MCP Server

An agent-facing experiences layer for Phuket that lets AI assistants (ChatGPT, Claude, etc.) answer "what should we do here, today, for this exact party" in one tool call — with safety-gated attributes, live weather/sea context, and affiliate booking links.

Live: thailandfunexperiences.com · MCP endpoint: https://thailandfunexperiences.com/mcp

What it does

• 1,891 real Viator experiences synced, 352 enriched with 24 safety/suitability attributes each
• Portfolio ranking — returns 4 diverse candidates (not top-k), with venue dedup, activity-type diversity, and category spread
• Live context — weather forecasts (slot-aware rain buckets), sea state (swell direction × zone exposure), travel matrix (ORS-routed, 13 zones), season
• Safety gating — age floors, mobility (enjoy-the-core semantics), pregnancy, non-swimmer, motion comfort, all with multi-source evidence hierarchy
• Affiliate booking — every result includes a tracked booking link (book_now_url) with live availability

Architecture

See ARCHITECTURE.md for the full spec.

Agents (ChatGPT / Claude / REST)
        │
        ▼
  M6 MCP Server (Fastify + streamable HTTP)
        │
  M5 Ranker (pure function, no DB/API calls)
   ├── M1 Experience Graph (Postgres + PostGIS)
   ├── M4 Context (Open-Meteo weather/marine, ORS matrix)
   └── M3 Enrichment (Claude Sonnet, Batch API)
        │
  M8 Router → Viator affiliate links
  M10 Telemetry → session/click/booking events

Tools

search_experiences

Search Phuket activities for a travel party. Returns a portfolio of ~4 diverse recommendations.

{
  "date": "2026-09-10",
  "staying": "kata",
  "party": [{"role": "adult", "age": 38}, {"role": "child", "age": 9}],
  "constraints": {"motion_comfort": "low"},
  "energy": "high",
  "time_slot": "morning",
  "max_results": 4
}

Response includes: candidates with attributes, booking links, reason codes, weather context, sea state, catalog breadth stats, and refine hints.

get_experience

Get full details for a specific experience by ID.

{"experience_id": "exp_27613P9"}

Stack

• Runtime: Node 22, TypeScript (strict), ESM
• Monorepo: pnpm workspaces
• DB: Postgres 16 + PostGIS, Drizzle ORM
• MCP: @modelcontextprotocol/sdk streamable HTTP transport behind Fastify
• Enrichment: Anthropic Batch API (Claude Sonnet 4), structured extraction with evidence validation
• Weather/Marine: Open-Meteo (no key)
• Travel matrix: OpenRouteService (one batched matrix call for all 13 zones)
• Deploy: systemd + Caddy reverse proxy on VPS

Evidence hierarchy

1. human_correction    — Dan's per-attribute overrides
2. operator_terms      — verified T&C from specific operators
3. viator_structured   — supplier-declared flags (stored verbatim)
4. dan-rules           — category/venue-level local knowledge (8 rule groups)
5. derived             — formula-computed (seasickness = vessel × water_exposure)
6. extraction          — LLM structural/textual inference (base)

Key design decisions

• Portfolio, not top-k — 4 diverse candidates from distinct venues/activity types
• Structured constraints, not free text — no health language in any field (privacy by design)
• Transfer penalty always on — transfer/duration ratio ≥1 = severe penalty
• Graceful decline for unsupported destinations — never throws, logs demand
• Directive 9 (Viator-tomorrow test) — every served feature must add transformation, corroboration, composition, or context a single provider can't replicate

Built with

Built solo in ~1 week with Claude Code (Claude Opus). See meta/POSTMORTEM.md for the build retrospective.

License

MIT