Discogs MCP Server

A remote Model Context Protocol server for Discogs, hosted on Cloudflare Workers. Two things it's good at:

1. Finding the best-sounding pressing of an album — surveys every version of a master release, scores each on community ratings, collector demand, format quality, and audiophile markers (MFSL, half-speed masters, renowned mastering engineers, Japanese pressings…), and returns a ranked comparison.

2. Mood- and taste-based recommendations — mood-aware search of your own collection ("mellow Sunday morning"), catalog discovery ranked against your taste profile, and cross-user discovery that mines other collectors' public collections.

Auth is browser-based Discogs OAuth: connect an MCP client to the server URL and you'll be redirected to Discogs to log in. Read-only — the server never modifies your collection.

Quick start (self-hosting)

Prerequisites: Node 18+, a Cloudflare account, and a Discogs developer app.

npm install
npm run setup:kv          # creates OAUTH_KV and CACHE_KV; paste the IDs into wrangler.toml
cp .dev.vars.example .dev.vars   # fill in your Discogs consumer key/secret
npm run dev               # local dev on http://localhost:8787

Register the Discogs app with callback URL http://localhost:8787/callback for dev. Test with the MCP Inspector:

npm run inspect

Deploy

wrangler secret put DISCOGS_CONSUMER_KEY
wrangler secret put DISCOGS_CONSUMER_SECRET
npm run deploy

Then update your Discogs app's callback URL to https://<your-worker>.workers.dev/callback.

Connect a client

• Claude Code: claude mcp add --transport http discogs https://<your-worker>.workers.dev/mcp

• Claude Desktop / claude.ai: add a custom connector with URL https://<your-worker>.workers.dev/mcp

• Legacy SSE clients can use /sse instead of /mcp.

To restrict who can log in, set ALLOWED_DISCOGS_USERS in wrangler.toml (comma-separated Discogs usernames and/or numeric user IDs).

Related MCP server: Navidrome-MCP

Tools

Prompts: find-best-pressing, recommend-by-mood, cross-user-discovery, compare-pressings. Resources: discogs://collection, discogs://wantlist, discogs://release/{id}, discogs://master/{id}, discogs://master/{id}/versions, discogs://user/{username}/collection, discogs://user/{username}/wants.

Example conversations

"What's the best pressing of OK Computer?" → find_best_pressing ranks the top versions with score breakdowns and audiophile signals.

"What should I listen to on a rainy evening — something I already own?" → search_collection maps "rainy" to Cool Jazz / Ambient / Shoegaze / Post-Rock / Trip Hop and filters your collection.

"Compare my collection with user xyz and tell me what to buy next." → discover_similar reports your profile similarity and their albums that best match your taste.

Architecture notes

• Auth bridge: MCP clients speak OAuth 2.1 to the Worker (via @cloudflare/workers-oauth-provider); the Worker speaks OAuth 1.0a to Discogs. Discogs tokens live encrypted inside the MCP access token (props) — no server-side session store.

• Caching: all Discogs reads go through a KV read-through cache (releases/masters 24 h, versions 12 h, search 6 h, collections/wantlists 4 h). Collections are additionally cached as a single slim aggregate, so mood search, stats, and recommendations cost zero API calls when warm.

• Rate limits: Discogs allows 60 req/min authenticated. The client retries 429s with exponential backoff and soft-throttles when the remaining budget is low. find_best_pressing budgets calls by pre-ranking versions on collector signals and fetching details only for the top ~12 candidates (the versions endpoint carries no ratings).

• Collections are fetched at 100 items/page up to 3,000 items; beyond that results are truncated and flagged (truncated: true).

Development

npm test        # vitest unit tests (scoring, mood mapping, similarity)
npm run lint    # tsc --noEmit
npm run build   # wrangler dry-run bundle

License

MIT