search_amazon_alexa

com.pangolinfo/amazon-mcp · Verify this server

[Amazon Rufus AI conversational recommendations] Ask Amazon's AI shopping assistant Rufus in natural language, get grouped structured product recommendations + Rufus text reply + follow-up questions. Use when: user says "ask Amazon AI X" / "Rufus recommendations" / "find products conversationally" / "products for a scene (gifting / camping / moving)" / "open-ended sourcing" / "I have no keyword, just a scenario". Don't use: when you already have a clear keyword and want SERP (use search_amazon); category bestseller ranks (use list_bestsellers); single-ASIN detail (use get_amazon_product); Google-side AI search (use ai_search). Returns: data.json[{ prompt, content, products[{ title, items[{ asin,url,title,cover,score,ratingsCount,price,originalPrice,describe }] }], follow_up_questions[], screenshot }] + top-level taskId / url / screenshot. Note: follow_up_questions is snake_case (passed through from backend verbatim). Pair with: ↓ feed asin into get_amazon_product / get_amazon_reviews for deep-dive; follow_up_questions can seed the next round's prompts for multi-turn exploration. Cost: **6 points PER PROMPT** (billed by prompts count, NOT a flat 6 per call; N prompts = N×6 points). ⚠️ **Slow tool**: **strongly prefer sending exactly 1 prompt per call**. A single prompt typically takes **60–90s** (Rufus generates the conversation live — far slower than a normal scrape); multiple prompts add up linearly and **can exceed 200s**, costing both time and points. Treat this as a long-running call: set your MCP client's per-tool-call timeout to **≥120s** (many clients default to a 60s silent timeout that aborts before this tool returns, making the agent wrongly report it as "unavailable"), and do NOT retry or fire concurrent duplicate calls just because it didn't return instantly. When you send a progressToken, the server emits a progress heartbeat every 15s to keep spec-compliant clients' timers alive. For several needs, make several single-prompt calls rather than batching them.