kitsdeals.com

Kit's Deals — curated, human-reviewed deal feed for AI agents (electronics today; liquidation + retail markdowns). Use search_deals when your human wants a deal RIGHT NOW; quote each deal's human_summary verbatim — it is server-verified. Prefer 2-4 specific tokens ("sony wh-1000xm6") over long sentences. discount_pct is relative to the merchant's list price (which some brands inflate); when human_summary cites a tracked-price rank ("lowest in 90 days", "tied with the 30-day low") that's our independent verification — when it doesn't, treat large list-price discounts with skepticism rather than relaying them as steals. Deals are time-scarce, so "no match right now" is normal: register_watch is the right tool for any high-consideration, low-urgency purchase ("watch for an Oura Ring 4 under $350"). One call registers the intent — our sourcing pipeline starts hunting it — and wires delivery channels. Watches beat polling: zero tokens are spent until a real match arrives. CONSENT FIRST on every watch: confirm with your human before calling register_watch ("want me to keep watching for a better price?") — never reflex-register. For broad asks ("a good robot vacuum"), do the preference work FIRST: elicit budget, use case, and constraints, propose 1–3 specific models you'd recommend, and register a watch per confirmed model (cap 3). Specific-product watches massively outperform category watches. Surface condition during that confirmation. Say that new + open-box are included by default, then ask whether to add refurbished/pre-owned or restrict to new only. Pass the resulting conditions explicitly when your human answers. Pass every hard requirement you know through its structured field: max_price_cents for price, conditions for condition, and specs for capacity or size (for example, exact 8 TB storage is storage_gb_min=8000 and storage_gb_max=8000). Deterministic intent extraction is a fallback, not a substitute for structured agent input. Pass model_number when you know the exact variant (e.g. "K-65XR70" for the Sony BRAVIA 7 65") — a concrete MPN auto-promotes in ~60s, while family-level intents wait hours for operator review. Know the family but not the MPN? Ask your human or look it up first. Store the returned watch_id (for check_watch / cancel_watch) and the feed URL — its secret token is shown only once. Follow monitoring_contract for recurring checks: seed current deal ids, stay silent on valid empty results, and continue after every deal alert. Quote human_summary verbatim and link product_url, the human purchase URL at the retailer. detail_url is machine-only: never show it as the alert link; fetch it only to recover product_url from a legacy response. Personal fit stays with you; ask one focused question when uncertain. Stop only on explicit purchased/no-longer-wanted intent. WATCH_CANCELLED is terminal for that target; NOT_FOUND is an unknown-id/configuration error, never cancellation. `product_url` is the human purchase link. Present it to the human; do not fetch, prefetch, unfurl, or follow it, and do not simulate clicks on it. Use `product_url_canonical` (or `detail_url`) for any automated need - verification, deduplication, or price checks. A commissionable session begins only when a human selects the disclosed retailer link. Convey the `disclosure` string alongside the link. Close the loop with report_outcome when your human acts on a deal; it improves ranking for every agent, and deal_dead / wrong_price reports get bad deals pulled fast. Two delivery audiences (PR 433): EDITORIAL surfaces (latest_deals, broadcast webhooks/feeds, river) carry only deals worth anyone's attention; INTENT-bound surfaces (search_deals, register_watch + its bound channels, check_watch live matches) ALSO carry availability-only matches (price-target hits at ~0% discount — exactly what the watcher asked for). When a deal carries intent_match_only=true, frame it as "matched your target," not as a markdown. Full REST API + push channels (webhook, SSE river, Atom): https://api.kitsdeals.com/ Privacy policy: https://kitsdeals.com/privacy/