ebay-finder

name	ebay-finder
description	Find anything on eBay from a vague, plain-language wish — in ANY language — even with fuzzy constraints like "must actually work", "best value for money", a date range ("1930–60"), or a vibe ("looks well-loved, not beat up"). Translates the wish into expert eBay searches, fans out across angles, inspects listings (incl. photos), and returns a ranked shortlist with reasons. Works with NO API key; richer with a free eBay developer key. Use whenever someone wants to buy, find, price-check, or hunt for something on eBay and plain keyword search isn't enough.
license	MIT

🔭 eBay Finder

Plain keyword search fails the moment a wish has soft constraints — "a 1930–60 typewriter, has to work, best value for money". eBay's own AI shopper is closed/US-only. This skill is the open one that works for everyone: you are the intelligence — you read the wish, design the search strategy, judge the photos and prices — and the bundled Python tool does the deterministic parts (build precise eBay URLs, optionally call the official Browse API, and score listings transparently).

When to use

• "Find me a …", "help me buy …", "what's a good … on eBay", "is this a fair price"
• Any wish with fuzzy criteria: condition/working state, era, budget, brand vibe, value-for-money, "not a project / not for parts", "with original box", etc.
• Requests in any language (translate the intent, keep the user's language in the report).

The method (follow in order)

1 — Understand the wish → Criteria

Read the request and extract the soft goal. Build a criteria object:

• raw_request: the user's exact words (any language).
• item_summary: a short English description.
• must_work: true if functional/tested matters.
• value_focused: true for "best value / cheap but good / bang for the buck".
• target_price_usd: their mental anchor if any (convert currency).
• era: e.g. "1930-1960". deal_breakers: phrases that disqualify (e.g. "repainted", "reproduction"). nice_to_haves: phrases that boost (e.g. "original case", "new ribbon", "serviced").
• notes: 1–2 sentences explaining your search strategy (shown in the report).

If the wish is genuinely ambiguous in a way that changes the search (budget unknown AND no anchor, or item type unclear), ask one crisp question. Otherwise proceed — don't interrogate the user.

2 — Strategize → several QueryPlans

Vague wishes need multiple search angles, not one query. Typical fan-out:

• A broad angle (category + era keywords).
• One angle per likely brand / synonym (e.g. typewriters → Royal, Remington, Underwood, Smith-Corona, Olympia). Brands are where the value hides.
• A budget angle (tight price band, sorted by price).
• A fresh angle (NEWLY_LISTED) to catch underpriced new listings.

For each plan set: label (why this angle), keywords, category_id (see reference/ebay-search-grammar.md), conditions, buying_options, price_min_usd/price_max_usd, sort, item_location_country, aspects (e.g. [["Brand", ["Royal","Remington"]]]).

Encode must_work as ranking intent, not a hard filter — sellers mislabel constantly. The tool adds conservative negative keywords (-"for parts") for you.

3 — Search → always end with real listings

The deliverable is a ranked list of actual eBay listings (each an ebay.com/itm/… link) that satisfy the user's inputs — NEVER a bare list of search-page URLs. Search-page URLs are a fallback for the human to keep browsing, shown in addition to listings, never instead of them. Walk down this ladder until you have listings:

First build the plans into URLs + a spec the tool understands:

python -m ebay_finder.cli plan.json --out report.md   # always: builds the search URLs

Discovery ladder — stop at the first rung that yields listings:

1. API mode — if EBAY_CLIENT_ID + EBAY_CLIENT_SECRET are set, the tool calls the official Browse API and fetches + dedupes + ranks structured listings automatically. (See SETUP.md for the free 2-minute key.) Best quality. Done.

2. Fetch the search URLs — open each generated /sch/ URL with your fetch/browser tool, read the listing cards, and collect items into listings.json.

3. eBay-restricted web search (universal fallback — works even when fetch is blocked). If you cannot fetch eBay pages (datacenter IPs are often challenged), use your web search tool restricted to ebay.com, one query per plan, e.g. Olympia SM3 typewriter serviced working with allowed_domains=["ebay.com"]. Keep the ebay.com/itm/<id> results — those are real, current listings. This rung ALWAYS produces item links, so you are never left with only search pages.

Then rank whatever you gathered:

python -m ebay_finder.cli plan.json --listings listings.json --out report.md

listings.json = a JSON array of items, each: title, url (the /itm/ link), and any of price_usd, shipping_usd, condition_raw, seller_feedback_pct, item_location, image_url. Missing fields are fine — rank on what you have and flag unknowns. If you only have titles + links (rung 3), still return them as a ranked shortlist using working/condition signals in the titles, with fair-price bands and a clear "confirm price ≤ budget" note.

4 — Inspect (use your eyes)

For the top candidates, look at the photos (image_url): confirm the era/model, spot cracks/rust/missing keys/repaints, sanity-check that "working" claims are plausible. Read the seller's condition text. This visual judgment is your job and is exactly what keyword search can't do — it's the skill's superpower.

5 — Rank & report

The ranker scores 0–100 on value-vs-cohort, working-confidence, condition fit, seller trust, returns, deal-breakers and nice-to-haves, with a reason on every line. Lead with the ranked list of actual listings — each with its ebay.com/itm/… link, total price (incl. shipping), why it ranked, and any ⚠ flags ("for-parts", "no-returns", "low-feedback", "price-outlier-low"). Add your own one-line verdict and a fair-price band. Only after the listings, you may add the raw search URLs so the user can keep browsing. Reply in the user's language. Never reply with only search-page URLs.

Plan spec example (flagship: the typewriter)

See examples/typewriter.json for a full, runnable spec. The shape:

{
  "criteria": {
    "raw_request": "1930-60 type writer, has to work, high value for money",
    "item_summary": "working vintage typewriter, 1930s–1960s, best value",
    "must_work": true, "value_focused": true,
    "target_price_usd": 120, "era": "1930-1960",
    "deal_breakers": ["repaint", "reproduction"],
    "nice_to_haves": ["serviced", "new ribbon", "original case"],
    "notes": "Fan out across the value brands; cap budget; reward tested/serviced."
  },
  "plans": [
       
          
            
         
          
  

Principles

• Always deliver listings. In every case — API, fetch, or eBay-restricted web-search fallback — return a ranked list of real ebay.com/itm/… listings that satisfy the user's inputs. Search-page links are a bonus for further browsing, never the answer.
• Respect eBay. Use the official Browse API when keyed; keyless = build URLs, do polite human-initiated fetches, or use web search restricted to ebay.com. Never build a high-volume scraper or automate the eBay UI — eBay's user agreement forbids agentic bots.
• Be honest. Prices/availability change fast; tell the user to verify before buying. Surface risk flags, don't hide them to make a result look good.
• Multi-language by default. Understand any language; answer in theirs.

Files

• ebay_finder/ — the tool (zero runtime deps, stdlib only).
• reference/ebay-search-grammar.md — eBay URL + Browse API cheat sheet.
• reference/discovery-ladder.md — how to always return real listings (API → fetch → eBay-restricted web search).
• examples/ — runnable plan specs.
• SETUP.md — optional free eBay API key in ~2 minutes.