@pipeworx/sothebys

Connect: https://gateway.pipeworx.io/sothebys/mcp · Install: one-click buttons

Tools: 2

Realized auction prices from Sotheby’s — search the past-lot archive by artist or maker across every sale, and get the price including buyer’s premium, the winning (hammer) bid, the estimate range and the sale’s name/date/location; pull full catalogue detail (description, provenance, literature, exhibition history, images) for one lot.

Tools

  • sothebys_results_search(query, category?, sale_type?, page?, limit?, include_unpriced?) — search the archive by artist/maker/keyword. category is a free-text filter matched against the lot’s department names (e.g. “prints”, “photographs”). sale_type is "timed" (default), "live" or "any" — see What is and isn’t public below, it is the single most important argument here.
  • sothebys_lot_details(lot_id) — full catalogue detail for one lot, keyed by the lot_id a sothebys_results_search result carries.

Tools are prefixed with the house name deliberately: three art-house packs (Bonhams, Sotheby’s, Phillips) expose the same two shapes, so bare results_search would collide across them and make ask_pipeworx routing ambiguous.

Auth

Keyless. Both upstreams serve anonymous callers; the Algolia search key is minted on demand by Sotheby’s own public GraphQL and cached for 5 minutes.

Data sources

Sotheby’s is not the Bonhams shape. Its lot pages are a Next.js app fed by a separate GraphQL backend, so there is no server-rendered search page to read. Two public, keyless endpoints carry everything, and neither is /bsp-api/*:

  • https://clientapi.prod.sothelabs.com/graphql — Sotheby’s own public bidding API. Unauthenticated POST works. Introspection is disabled by its Cosmo Router, but the operation documents ship in the site’s own JS bundles under /buy/_next/static/chunks/. Three operations are used:
    • AlgoliaSearchKeyQuery → a search key for the index below. Requesting no filters yields a key scoped to the whole index, not one auction.
    • LotCardsQuery(id, lotIds, language)auction.lotCardsFromIds, the per-auction batch price lookup. bidState.currentBidV2 is the winning (hammer) bid; bidState.sold.premiums.finalPriceV2 is the realized price including buyer’s premium. Amounts arrive as decimal strings. auction.testRecord flags Sotheby’s QA sales, which are in the index.
    • lotV2(lotId, language, countryOfOrigin) → full catalogue detail. It returns the Lottable interface, so the selection needs ... on LotV2 { … }; condition is an object, not a string.
  • https://KAR1UEUPJD-dsn.algolia.net/1/indexes/prod_lots/query — the search index the site itself queries (window.ALGOLIA_APP_ID / window.ALGOLIA_INDEX_NAME in the page source). This is the only cross-auction search path. It carries creators, departments, estimates, auctionType, auctionDate and the site-relative lot slug — but no prices at all: price is null on every record, which is why every search costs a second, per-auction GraphQL round trip.

Neither host is www.sothebys.com, so the Crawl-delay: 15 and the Disallow: /bsp-api/* in https://www.sothebys.com/robots.txt (checked 2026-09-05) do not cover them; clientapi.prod.sothelabs.com serves no robots.txt at all (404). This pack makes no request to www.sothebys.com — it only builds www URLs for a human to click.

What is and isn’t public

The realized price comes from the GraphQL bidState.sold union, which is either ResultVisible or ResultHidden depending on whether Sotheby’s publishes that sale’s results through this platform. Measured 2026-09-05 across the 44 distinct sales in a 100-hit “Pablo Picasso” sample, the split is a clean rule rather than noise:

slicesales visible
auctionType: Live (saleroom sales), any year including 2025–260 of 11
auctionType: Timed (online sales), ~2021 onwards18 of 18, every lot
auctionType: Timed, before ~20210 of 15

So sale_type defaults to timed: that is the slice of the archive where a realized price exists at all. Asking for live returns lots with null prices however famous they are — Sotheby’s publishes live-saleroom results on its own results pages, not through this API — and the tool says so in note rather than letting it read as a parse failure.

Two consequences worth knowing when reading a response:

  • sothebys_results_search scans a page of up to 100 search hits and returns the priced subset, so count is normally well below scanned, and total_hits (the whole archive match count) is far larger than both.
  • Price lookups are spent newest sale first, not in relevance order. Relevance happily ranks 2019 sales top for “Pablo Picasso”, and those can only answer “hidden”; output order is still relevance order.
  • bonhams — same two-tool shape for Bonhams, whose realized prices are plain server-rendered HTML.
  • Phillips is the third house in this survey; Christie’s is closed (internal API) and out of scope.

Tools

  • sothebys_results_search — Search Sotheby’s past-lot archive by artist/maker (or any keyword) and return realized prices: price including buyer’s premium, winning (hammer) bid, estimate range, sale name/date/location and depart
  • sothebys_lot_details — Full detail for one Sotheby’s lot: catalogue description, provenance, literature, exhibition history, catalogue note, images, estimate range, and the realized price (including buyer’s premium) where S

Tools

  • sothebys_lot_details — Full detail for one Sotheby's lot: catalogue description, provenance, literature, exhibition history, catalogue note, images, estimate range, and the realized price (including buyer's premium) where S
  • sothebys_results_search — Search Sotheby's past-lot archive by artist/maker (or any keyword) and return realized prices: price including buyer's premium, winning (hammer) bid, estimate range, sale name/date/location and depart

Regenerated from source · build September 5, 2026