📚 goodreads-mcp

A read-only MCP server for Goodreads — built without the Goodreads API, because there hasn't been one since December 2020. Lets an LLM find and research books, ratings, and reviews. Tools ride on RSS feeds, the JSON autocomplete endpoint, and the __NEXT_DATA__ blob embedded in book pages. No login, no cookies, no writes — public data only.

tools

The discovery tools all take a book_id and return results carrying book_id/title/author/rating/url, so an agent can chain them — e.g. similar_books → get_reviews on a recommendation. This is the structured book graph a general web search can't assemble.

WAF note: Goodreads book HTML pages now sit behind an AWS WAF JavaScript challenge (HTTP 202) that plain HTTP clients can't solve. get_book routes around it via the .xml-suffixed page, so it still works without a browser. If Goodreads ever extends the WAF to a path we depend on, the client raises WAFChallenge with a clear message instead of a confusing parse error.

Related MCP server: zotero-mcp

install

cd goodreads-mcp
python3.10 -m venv .venv && .venv/bin/pip install -e .

Requires Python ≥ 3.10.

config (optional)

No login or cookies — everything is public data. The only setting is your numeric user_id, the default for the shelf tools. It's the number in goodreads.com/user/show/<ID>-yourname; you can also pass user_id to each shelf tool per call.

mkdir -p ~/.config/goodreads-mcp
cat > ~/.config/goodreads-mcp/config.json << 'EOF'
{ "user_id": "12345678" }
EOF

Env var GOODREADS_USER_ID overrides the file.

Claude Desktop config

~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "goodreads": {
      "command": "/path/to/goodreads-mcp/.venv/bin/goodreads-mcp"
    }
  }
}

Or for development, mcp dev goodreads_mcp/server.py gives you the Inspector UI to poke each tool.

first-run verification

The endpoints are unofficial, so verify in this order:

1. search_books("project hail mary") — should just work

2. get_book("54493401") — confirms the .xml/WAF workaround; check the histogram is populated

3. get_reviews("54493401") — should return real review text

4. get_shelf("to-read") — checks your user_id + RSS

5. list_shelves() — best-effort shelf-name scrape

tests

.venv/bin/pip install -e ".[test]"
.venv/bin/pytest                       # offline parser/unit tests
GOODREADS_LIVE=1 .venv/bin/pytest      # + live network smoke tests

design notes

• Request-first, no browser automation. Everything is httpx against JSON/RSS/embedded-JSON/GraphQL surfaces; the only HTML regex is in list_shelves and the GraphQL config discovery.

• GraphQL backbone (reviews). get_reviews calls Goodreads' AppSync GraphQL endpoint — the same backend the website uses. The web app ships a public read-only API key in its JS bundle; the client scrapes the endpoint + key from that bundle at runtime and caches them, so a key rotation self-heals (client.graphql_config). A hardcoded pair is kept as a fallback. This is what enables real pagination (past the ~30 reviews a page embeds) and server-side rating filters. GraphQL partial-success is respected: a deleted review's sub-resource just comes back null rather than failing the call.

• WAF-aware. Book pages sit behind an AWS WAF JS challenge; get_book uses the .xml path that isn't gated, and the client raises WAFChallenge if it ever gets a challenge body so failures are loud, not silent. (The GraphQL endpoint is a separate AppSync host and isn't WAF-gated.)

• Polite client. Single persistent session, browser-faithful headers, exponential backoff on 429/503; get_reviews caps paging at 100 reviews.

• Caveats: all of this is unofficial and depends on markup/endpoints/keys that can drift.

ideas for v2

• richer book data: series, similar books, top reviews from the Apollo state

• author pages / bibliography

• caching layer for repeated lookups