Deal-Hunter + Auto-Bid Agent — Design Scaffold¶

Goal: find the best value-for-fit used/new RAM (and later: any hardware) across EU marketplaces, rank it, and bid/buy within a budget — compatibility-aware. Date: 2026-06-24. Status: design only (to implement/iterate).

1. Does this already exist? (buy-vs-build)¶

No off-the-shelf product does the whole thing. Every layer exists as a building block; the connective tissue is the custom value.

Layer	Off-the-shelf?	Use / Build
Auction bid mechanics (auto max-bid snipe)	✅ Solved — Gixen (free, ~$12/yr Mirror, HTTP API, OAuth, EU); eBay native max-bid as fallback	REUSE — never build bidding
eBay discovery	✅ Official Browse API (free, OAuth client-creds, ~5k calls/day)	REUSE
Per-site price alerts	✅ Keepa (Amazon, API), Geizhals/Idealo/Tweakers "Preisalarm", Pepper (API+RSS)	REUSE
Per-site classifieds alerts	✅ Native saved-search EMAIL alerts on Marktplaats / 2dehands / Kleinanzeigen / Leboncoin (free)	REUSE via Gmail MCP
Multi-site classifieds monitor	✅ OSS: `BoPeng/ai-marketplace-monitor` (LLM-scored, Claude-capable), `lbc-finder`, `marktplaats-monitor`, `GoddeerisEdouard/MarketplaceNotifier` (BE!); SaaS: CollectAlert €2.99/mo, Kleinanzeigen-Agent API	REUSE/FORK
Payment with budget caps	✅ Google AP2 Intent Mandate, Visa Intelligent Commerce, Stripe/PayPal agent toolkits	REUSE if ever needed (not needed for v1–2)
Cross-lane aggregation (new retail + used classifieds → one ranked, all-in-cost list)	❌ Gap	BUILD
Hardware-compatibility-aware filtering ("DDR4-3200 1.2 V SO-DIMM that fits my laptop")	❌ Gap — every tool filters on keywords+price only	BUILD (the core value)
Decision → bid (connect a discovered underpriced auction to an actual bid within budget)	❌ Gap	BUILD (orchestration)

Verdict: Hybrid. Reuse Gixen (bidding) + eBay Browse API + native saved-search emails + an OSS monitor for sourcing. Build only: (1) compatibility-aware extraction/filter, (2) cross-platform value scoring, (3) the orchestration that turns a scored deal into a snipe.

2. ⚠️ Legal / ToS constraints (these shape "full-auto")¶

eBay bans third-party autonomous LLM "buy-for-me" / checkout agents from 2026-02-20 (allowlisted partners only). BUT sniping with a human/agent-preset max bid is explicitly allowed. → The compliant "full-auto" path is: agent computes a max bid in advance from valuation → pushes it to Gixen → Gixen snipes at close. That's scheduled automation (allowed), not live LLM decide-at-close (banned). Stay on this side of the line.
No open public bid API — eBay's Offer placeProxyBid is partner-gated "Limited Release." So programmatic bidding = via Gixen's API, not a DIY eBay bid call.
Classifieds have NO buyer/checkout API and prohibit scraping; there's no auction (fixed price). → "Full-auto purchase" is not compliant there. Best you can do: auto-discover + auto-rank + (optional) auto-draft a seller inquiry; a human completes payment/pickup.
Facebook Marketplace: no buyer API → exclude, or alert-only with human in loop.

So "full-auto" in practice = auto-bid on eBay auctions up to a budget-capped, pre-computed max (via Gixen); everything else is auto-rank + alert/▶human. Always: global budget cap, per-item max, dry-run mode, kill-switch, full audit log.

3. Architecture — 5-stage pipeline¶

WATCH ─▶ EXTRACT ─▶ VERIFY(fit) ─▶ SCORE(value) ─▶ DECIDE ─▶ ACT

| Stage | Tech (reuse) | Output | |---|---|---| | Watch | eBay Browse API (auctions+fixed) · classifieds saved-search emails (Gmail MCP) · ai-marketplace-monitor fork · optional CollectAlert/Kleinanzeigen-Agent | raw listings stream | | Extract | Claude (Anthropic API) parses title/desc/photos → schema | structured listing | | Verify | rules + Claude vs. machine spec (from pc-hardware-ram memory) | PASS / FAIL / ASK-SELLER | | Score | value function (below) | deal_score, €/GB, suggested_max_bid | | Decide | threshold + budget cap | BID / ALERT / SKIP | | Act | Gixen API (snipe) · Gmail (digest / seller draft) | snipe placed / email sent |

Listing schema (Extract → JSON)¶

{ "source":"ebay.be", "url":"...", "title":"...", "type":"auction|fixed",
  "capacity_gb":64, "modules":2, "module_gb":32, "form_factor":"SODIMM",
  "speed_mts":3200, "voltage":1.2, "xmp":false, "ecc":false, "rank":"2Rx8",
  "part_no":"CT2K32G4SFD832A", "condition":"used-tested",
  "price_eur":180, "shipping_eur":7, "ships_to_be":true,
  "seller_rating":99.2, "buyer_protection":true, "photo_of_label":true,
  "ends_at":"2026-06-27T20:11:00Z" }

Compatibility filter (the BUILD core) — hard gates from machine spec¶

PASS requires: form_factor==SODIMM AND speed_mts>=3200 AND voltage==1.2
               AND xmp==false AND ecc==false
               AND (modules,module_gb) in {(2,16),(2,32)}
REJECT: DIMM(desktop), DDR4L/LPDDR4, 1.35V/XMP-only, ECC/RDIMM, speed<3200
ASK-SELLER if voltage/rank/tested unknown  →  queue a templated question

(Machine spec source: ASUS ROG G513QR, 2× SO-DIMM, 64 GB max, 1.2 V JEDEC DDR4-3200 — see pc-hardware-ram memory.)

Value function (the "best fit for value")¶

all_in = price_eur + shipping_eur + customs(if non-EU)
market_ref = baseline[capacity]      # used median: 32GB=€100, 64GB=€230 (configurable)
deal = (market_ref - all_in) / market_ref           # % under market
risk = w1*seller_norm + w2*tested + w3*photo + w4*matched + w5*buyer_protection   # 0.5–1.0
adjusted = deal * risk
suggested_max_bid = min( market_ref * TARGET_RATIO(=0.80), BUDGET_CAP )   # auctions
DECISION: BID if type==auction AND fit==PASS AND adjusted>=BID_THRESHOLD
          ALERT if fit==PASS AND adjusted>=ALERT_THRESHOLD
          else SKIP

4. Phased roadmap¶

v1 — Alert-only digest (build first, zero risk). eBay Browse API + Gmail-parsed saved-search alerts → Claude extract+fit+score → daily ranked email (via Gmail MCP) with deal scores + suggested max bids. No bidding.
v2 — eBay auto-snipe. Add Gixen API: for BID-decisions, place a budget-capped snipe (pre-computed max). Per-item approval toggle; hard global cap; dry-run + audit log. (Compliant: sniping allowed.)
v3 — Classifieds assist. Auto-draft seller inquiry (templated, incl. the ASK-SELLER spec questions) for top classifieds hits; human sends/pays. Optional MemTest86 reminder on receipt.
v4 — Generalize. Swap the fixed RAM spec for a pluggable "compatibility profile" so it works for SSDs, GPUs, any fitment-gated part.

5. Reusable components (fork/subscribe)¶

Bidding: Gixen (Mirror + API) — github refs: cyberfox/gixen, dgunthor/gixen4j.
eBay discovery: timotheus/ebaysdk-python or raw Browse API.
Classifieds sourcing: fork BoPeng/ai-marketplace-monitor (LLM-scored, Claude-ready) + GoddeerisEdouard/MarketplaceNotifier (BE) + etienne-hd/lbc-finder (FR).
Brain: Anthropic API (Claude) for Extract + Verify + Score. Alerts in/out: existing Gmail MCP.

6. Open decisions for implementation¶

Budget cap value + per-item max; ALERT vs BID thresholds; TARGET_RATIO.
Sold-comp data for market_ref (eBay Marketplace Insights is gated → start with our used/new baselines, refine from observed sales).
Run host: cron/Task Scheduler vs a small always-on service (needs to fire snipes near auction close — Gixen handles the actual last-second timing, so the agent only needs to register snipes ahead of time → cron is fine).
Storage: SQLite for listings/dedup/audit log.