calibreweb-mcp

A read-only Model Context Protocol (MCP) server for Calibre-Web (and Calibre-Web Automated), the self-hosted ebook library web UI.

Lets MCP clients like Claude Code, Claude Desktop or Codex search your library, browse the curated views and shelves, follow per-format download links and pull cover images straight into the conversation. It never writes anything: every tool is a GET.

Six tools is the ceiling, not the floor: CALIBRE_WEB_ALLOW_TOOLS=essential registers a curated five instead, and a model picks the right tool far more reliably from five than from six — see choosing which tools load.

Calibre-Web has no REST API — its only stable machine-readable interface is the OPDS catalog feed it serves for e-reader apps. This server speaks that feed: Atom XML with HTTP Basic auth in, structured book data out.

What makes it different

The API Calibre-Web never had. Calibre-Web exposes no REST API — its only stable machine interface is the OPDS Atom feed built for e-reader apps. These tools parse that feed into structured book data with numeric ids, per-format download URLs and bounded summaries.

Read-only by construction. All six tools are GETs. Redirects are refused so Basic credentials never travel, XML carrying a DOCTYPE is rejected outright, hrefs are locked to the configured origin, and metadata is marked as the untrusted data it is.

Requirements

• Node.js 22 or newer
• A Calibre-Web instance (developed against the current linuxserver/calibre-web image; Calibre-Web Automated works the same way)
• A Calibre-Web user for the server. The OPDS feed authenticates with the normal web login — use a dedicated account with only the View and Download roles, not your admin account. If the instance allows anonymous browsing, the server can also run without credentials.

Configuration

¹ Leave both unset for an instance that allows anonymous browsing; setting only one of them is a configuration error.

Choosing which tools load

CALIBRE_WEB_ALLOW_TOOLS and CALIBRE_WEB_DENY_TOOLS take comma-separated tool names; a trailing * matches a whole family. essential is a curated preset of five: search_books, list_books, list_shelves, get_shelf_books, get_stats.

CALIBRE_WEB_ALLOW_TOOLS=essential
CALIBRE_WEB_ALLOW_TOOLS=search_books,list_shelves
CALIBRE_WEB_DENY_TOOLS=get_cover

An entry that matches no tool aborts startup and names it, so a typo cannot silently hide a tool — an absent tool is not something anyone traces back to an environment variable. A filtered tool is never registered, so it is absent from tools/list and unknown to tools/call alike.

If you run several of these servers at once, mcp-hub is the other answer — its /hub endpoint replaces every server's tools with six meta-tools.

Installation

Claude Code

claude mcp add calibreweb \
  -e CALIBRE_WEB_URL=https://books.example.com \
  -e CALIBRE_WEB_USERNAME=reader \
  -e CALIBRE_WEB_PASSWORD=... \
  -- npx calibreweb-mcp

Claude Desktop

{
  "mcpServers": {
    "calibreweb": {
      "command": "npx",
      "args": ["calibreweb-mcp"],
      "env": {
        "CALIBRE_WEB_URL": "https://books.example.com",
        "CALIBRE_WEB_USERNAME": "reader",
        "CALIBRE_WEB_PASSWORD": "..."
      }
    }
  }
}

Codex

[mcp_servers.calibreweb]
command = "npx"
args = ["calibreweb-mcp"]
env = { CALIBRE_WEB_URL = "https://books.example.com", CALIBRE_WEB_USERNAME = "reader", CALIBRE_WEB_PASSWORD = "..." }

Docker

docker run -i --rm \
  -e CALIBRE_WEB_URL=https://books.example.com \
  -e CALIBRE_WEB_USERNAME=reader \
  -e CALIBRE_WEB_PASSWORD=... \
  ghcr.io/ni-c/calibreweb-mcp

Through mcp-hub

A client that cannot spawn a local process — ChatGPT connectors, Claude on the web, Cursor, LibreChat — reaches calibreweb-mcp through mcp-hub: one container serves many stdio MCP servers over Streamable HTTP, with an OAuth 2.1 login behind a single password and long-lived tokens for the clients that cannot do OAuth. Its /hub endpoint puts every server behind six meta-tools, so one connector reaches all of them without N×tool schemas in the model's context, and it speaks both protocol revisions — a question this server asks travels through it to the person at the far end.

Its /config/mcp.json uses Claude Code's format, so the entry is the one you already have:

{
  "mcpServers": {
    "calibreweb": {
      "command": "npx",
      "args": ["-y", "calibreweb-mcp"],
      "env": { "CALIBRE_WEB_ALLOW_TOOLS": "essential" },
      "denyTools": ["get_cover"]
    }
  }
}

allowTools and denyTools there are the hub's own per-server filter, which is not the same thing as *_ALLOW_TOOLS in env — the difference, and the mistake it invites, are in the client guide.

Tools

All tools are read-only (readOnlyHint: true).

Book entries include authors, tags, series (with index), rating, a bounded summary, a cover URL and per-format download URLs — ready-made links a human can open, since the model itself has no reason to download an EPUB.

Structured output

Every tool declares an outputSchema and answers with structuredContent alongside the text block, so a client can use the result without parsing prose:

{
  "untrusted": true,
  "source": "calibre-web",
  "totalFound": 2,
  "truncated": false,
  "books": [{ "id": 7, "title": "Dune", "authors": ["Frank Herbert"] }],
  "notes": ["Book titles, authors, tags, series and summaries come from …"],
}

The untrusted marker is a field and not only a line in notes, because a client that reads the structured half would otherwise have to find the warning in a list of sentences. The two tools without it are get_stats, which is four counters this server has checked are numbers, and get_cover, which reports an id, a media type from a four-entry allowlist and a byte count — the image itself stays in the content block where a client renders it.

An over-budget result drops book summaries as before. Where that is still not enough it is now an error rather than JSON cut at the ceiling: unparseable text was tolerable in a text block and is not something structuredContent can carry, and the two channels have to hold the same value.

Pagination

Feeds are paginated by the instance's books per page setting (default 60); the page size is not client-controllable. Every listing returns pagination.nextOffset when more pages exist — pass it as offset in the next call. The discover view is a random selection and not paginated.

Not exposed, on purpose

• No writes. The OPDS feed has none, and this server would not add any.
• No file downloads. Tools return download URLs, not ebook payloads.
• No facet browsing (authors/series/tags/publishers/languages/formats as their own index feeds). search_books covers those lookups; the routes exist and tools for them can be added if there is a real use case.

Safety

• The server is read-only by construction — GET requests only, no state anywhere.
• Book metadata is untrusted third-party data; every result says so, control characters are stripped, and XML entity processing is disabled (documents declaring a DOCTYPE or entities are refused outright).
• Responses are bounded before parsing (8 MB feeds, 1 MB covers) and again before they reach the model (per-book and per-response budgets).
• Feed hrefs are only passed through when they resolve to the configured origin over http(s) — a hostile feed cannot plant javascript:, file: or cross-origin URLs into the results.
• Redirects are refused so the Basic credentials can never be replayed to another host; covers are only passed through for real image content types.
• The password is scrubbed from the process environment at startup, and URLs are credential-redacted before they appear in any log or result.

Documentation

The full guide, tool reference and security notes live at calibreweb-mcp.ni-c.de (source in docs/).

Development

npm install
npm test            # unit tests against a stubbed OPDS feed, no instance needed
npm run lint
npm run build

See CONTRIBUTING.md.

Releasing

1. Update CHANGELOG.md and bump the version in package.json (+ lockfile).
2. npm run lint && npm run test:coverage && npm run build
3. Tag the release: git tag -s vX.Y.Z -m "vX.Y.Z" && git push origin vX.Y.Z

Contributing

Issues, discussions and pull requests are welcome — see CONTRIBUTING.md. For vulnerabilities please use private reporting rather than a public issue; the policy is in SECURITY.md.

License

MIT © Willi Thiel