search_activities

🎯 RENDERING DIRECTIVE — READ FIRST. When this tool returns 2+ activities, you MUST present them as an interactive HTML/React artifact (Artifacts block, not flowing prose). Each activity is a visually distinct CARD with a prominent "Find on Tripadvisor" button styled as an HTML button, NOT an inline markdown link. Include the activity-type badge (Sight / Experience) prominently. If the call is part of a larger trip plan, the plan itself should be an artifact containing the activity cards. Single-result responses may use prose.

Search Tripadvisor's "Things to Do" (sights + bookable experiences/tours) for a location, optionally filtered by free-text query, place type, and minimum rating.

DISTINCT FROM search_events: activities are ongoing (visit a museum, take a cooking class anytime); events are date-specific (a concert on June 21). Use this for "what should I do in X"; use search_events for "what's happening while I'm there".

USE THIS TOOL WHEN:

• The user asks "what should I do in X" / "things to do in X" / "tours in X"

• They name an activity type ("cooking classes", "boat tours", "museums", "wine tasting")

• They want recommendations based on their preferences

Inputs:

• location (string, required) — free-text city or neighborhood. Combined with query into a single Tripadvisor search.

• query (string, optional) — free-text filter on activity type. Natural language works: "cooking class", "boat tours", "wine tasting", "free walking tour".

• place_type_filter (enum, optional, default "both") — one of "sights" (free attractions like museums, viewpoints), "experiences" (bookable tours), or "both" (default).

• min_rating (float, optional) — minimum review score 0.0-5.0. Results without a rating are excluded when this is set.

• max_results (int, optional, default 15) — 1-50.

Returns ActivityOffer entries each with:

• offer_id — Tripadvisor's place_id, stable per activity. Use this to drill in via get_activity_details (when implemented).

• name — activity name.

• activity_type — "sight" (free, non-bookable) or "experience" (bookable tour).

• rating, review_count — 0-5 scale (Tripadvisor's native).

• description — short prose (often missing on generic city searches; usually present on specific-activity searches).

• location — text "City, Country".

• thumbnail — URL (NOT hotlink-safe — don't render as a photo element).

• highlighted_review — {text, mention_count} — a relevant review snippet.

• booking_url — Tripadvisor listing URL. For experiences, this is the path to Viator tickets; for sights, it's the info page.

No coordinates and no price. Tripadvisor's search endpoint surfaces neither. Use get_activity_details(offer_id) (when implemented) to get price + duration + a direct Viator URL for bookable experiences.

PRE-CALL ELICITATION — three branches:

Branch 1: User names a specific activity type. "Find cooking classes in Lisbon." → query="cooking class". Search immediately.

Branch 2: User asks for a recommendation. "What should I do in Lisbon?" — before searching, infer the user's interests from conversation context + your own memory of them ("they love food and wine", "they're into history"). Bake the interest into query. NOTE: The MCP tool does NOT read Claude's memory — you (Claude) do the inference and pass the resulting query string. If memory yields nothing actionable, fall to Branch 3.

Branch 3: User is vague and you have no preference signal. "Things to do in Lisbon?" with empty conversation context. Ask ONE clarifying question: "Any particular interest — food, history, outdoors, nightlife?" Then search.

RESULT PRESENTATION:

• Card-based artifact, one card per result.

• For Branch 2 (memory-driven), preamble: "Based on your interest in food and wine, here are top-rated experiences in Lisbon." — makes the inference legible.

• Card content: name, activity_type badge (Sight / Experience), rating + review_count, location, the highlighted_review.text as a 1-line testimonial, "Find on Tripadvisor" button → booking_url.

• Do NOT render thumbnail as a photo element (Tripadvisor's CDN hotlink-protects). Same no-photos rule as stays/events.

• For a single result, prose is fine.