Discogs Search Scraper

Searches Discogs, the largest music release database and marketplace — across releases (specific pressings), master releases (the canonical work), artists and labels.

For releases, returns the full tracklist with per-track credits and durations, catalog number, community rating, and the live marketplace price range.

What you get

One row per search hit, plus a SEARCH_SUMMARY row per query.

From search (always) — raw passthrough, shape varies by entityType

Discogs' own clean GraphQL object, spread verbatim, so the fields differ by type. For entityType: "release": discogsId, title, released, country, siteUrl, labels[], formats[], images, primaryArtists[]. Every row carries __typename so you can tell Release/MasterRelease/Artist/Label apart in one dataset.

From the release page (includeReleaseDetails, release only, on by default)

tracklist[] — each track's position, title, credits (remix/featuring artist, when present) and duration — plus, from the page's Schema.org JSON-LD: genre, catalogNumber, recordLabel, releaseOf (the master release + primary artist), aggregateRating (community score and count), and offers (live marketplace offerCount / lowPrice / highPrice).

Input

{
  "queries": ["radiohead"],
  "entityType": "release",
  "genre": "Rock",
  "format": "Vinyl",
  "maxItems": 50,
  "includeReleaseDetails": true,
  "surface": "auto",          // website first, official API fallback (default)
  "discogsToken": ""          // optional personal access token: 60 API req/min instead of 25
}

format, genre, style, country, year and decade all genuinely narrow the result count and compose with AND (measured: genre=Rock (5,964) × format=Vinyl (826) → 611 together). format/genre/style need Discogs' own exact facet text — "Trip Hop", not "triphop".

A search term is optional. Leave queries empty and set a filter to browse by facet alone — e.g. every Reggae release in the database.

entityType: four genuinely different shapes

Verified at the same query (radiohead) these carry genuinely different fields, not the same index re-labelled — master/artist/label results carry no discogsId at all, only a relay id.

Two data surfaces (surface)

Discogs' website sits behind a Cloudflare challenge that, from cloud egress, blocks a random share of requests rather than a fixed list of clients. So the actor has two ways to read the same index and tells you which one it used in every row's _source:

The two shapes are both raw passthrough of what Discogs returns, so they differ: website rows carry __typename, discogsId, siteUrl, labels[], formats[], primaryArtists[]; API rows carry type, id, uri, label[], format[], catno, genre[], style[], community.want/have. Release detail from the website attaches musicReleaseJsonLd (with offers low/high price) and a parsed tracklist (position, title, credits, duration); from the API it attaches releaseApi (the full release object: lowest_price, num_for_sale, community.rating, genres, styles, identifiers, images, notes) and the API's own tracklist (position, title, duration, extraartists[]). Every RESULT row declares all three columns (musicReleaseJsonLd, releaseApi, tracklist), unused ones null.

Known limits

API mode is rate-limited. 25 requests/min per exit IP without a token — one search page (100 results) plus one call per release detail — so 50 releases with details take about two minutes on the API; with a personal access token (discogsToken, generated at discogs.com/settings/developers) it is 60/min. A 429 is waited out (20 s) on a fresh exit IP, up to 4 times.

decade on the API is a year range. The API has no decade filter, so decade: "1990" is sent as year=1990-1999 there (honest: 1,914 vs the website's 1,899 for the same query). An explicit year wins.

A query is never half-and-half. If the website fails mid-query, the whole query restarts on the API from page 1 (the surfaces have different page sizes and shapes), so the pages already fetched are discarded, never mixed.

Detail is release-only. includeReleaseDetails fetches the tracklist and JSON-LD only for entityType: "release". It's a no-op (logged, not an error) for master/artist/label — those detail pages weren't built out for this version.

Filter values are exact-match, not fuzzy. A typo doesn't silently return everything — Discogs is unusually honest here — but it does mean

format: "vinyl"

Pagination is genuinely honest — 25 results/page, and past the real last page it returns an empty array rather than clamping back to page 1 — so an unlimited crawl (maxItems: 0) terminates correctly, but a broad query can mean tens of thousands of results.

Anti-bot posture

www.discogs.com is behind a Cloudflare managed challenge (

cf-mitigated: challenge

primp safari_26

Policy

robots.txt disallows marketplace/account/user-scoped paths (*/marketplace, */sell/*, /users/, /login, /mycollection, etc.) — none of which this actor touches. /search itself is not disallowed for a generic crawler (only a Sogou web spider group disallows it specifically). No ClaudeBot/anthropic-ai group. api.discogs.com/robots.txt is the same file, and the API is Discogs' documented public interface (unauthenticated access allowed with a User-Agent; a token only raises the quota).