audiobookshelf-mcp

A Model Context Protocol (MCP) server for Audiobookshelf, the self-hosted audiobook and podcast server.

Lets MCP clients like Claude Code, Claude Desktop or Codex browse your libraries, answer questions about what you own and what you have listened to, and — unless you switch it off — keep your listening progress, bookmarks, collections and playlists up to date: 44 tools, 29 read and 15 write.

Forty-four tools is the ceiling, not the floor: AUDIOBOOKSHELF_ALLOW_TOOLS=essential registers a curated eight instead, and a model picks the right tool far more reliably from eight than from forty-four — see choosing which tools load.

What makes it different

Every response is a projection, not the raw object. An expanded library item carries every audio file, track and chapter with full ffprobe metadata. The media tools answer with a compact shape instead, and detail="full" is there for when the raw object really is what you want.

Twenty-nine of the forty-four tools only read. AUDIOBOOKSHELF_READ_ONLY=true registers those and nothing else, so a write tool is absent from tools/list rather than refused when it is called.

The six tools that take something out ask a person first, through MCP elicitation — a dialog the model cannot answer on its behalf, falling back to a single-use token bound to the exact targets where the client cannot show one.

Requirements

• Node.js 22 or newer
• Audiobookshelf 2.26.0 or newer — earlier versions have no API keys
• An Audiobookshelf API key

Getting an API key

API keys are managed by an admin under Settings → Users → API Keys. A key acts on behalf of exactly one Audiobookshelf user and inherits that user's permissions, so a key issued for a normal account cannot see libraries that account cannot see, and cannot delete anything unless that account may delete. The key is shown only once, at creation.

Configuration

The server starts without configuration: it completes the MCP handshake and lists its tools, and every call then fails with the setup instructions. That is deliberate, so registries and sandbox inspectors can introspect it.

Choosing which tools load

AUDIOBOOKSHELF_ALLOW_TOOLS and AUDIOBOOKSHELF_DENY_TOOLS take comma-separated tool names; a trailing * matches a whole family. essential is a curated preset of eight: list_libraries, search_library, list_library_items, get_library_item, get_item_chapters, list_items_in_progress, get_media_progress, set_media_progress.

AUDIOBOOKSHELF_ALLOW_TOOLS=essential
AUDIOBOOKSHELF_ALLOW_TOOLS=search_library,get_library_item,set_media_progress
AUDIOBOOKSHELF_DENY_TOOLS=delete_*

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, exactly like a write tool under AUDIOBOOKSHELF_READ_ONLY.

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 audiobookshelf \
  -e AUDIOBOOKSHELF_URL=https://abs.example.com \
  -e AUDIOBOOKSHELF_API_KEY=… \
  -- npx -y audiobookshelf-mcp

Claude Desktop

claude_desktop_config.json:

{
  "mcpServers": {
    "audiobookshelf": {
      "command": "npx",
      "args": ["-y", "audiobookshelf-mcp"],
      "env": {
        "AUDIOBOOKSHELF_URL": "https://abs.example.com",
        "AUDIOBOOKSHELF_API_KEY": "…"
      }
    }
  }
}

Codex

~/.codex/config.toml:

[mcp_servers.audiobookshelf]
command = "npx"
args = ["-y", "audiobookshelf-mcp"]
env = { AUDIOBOOKSHELF_URL = "https://abs.example.com", AUDIOBOOKSHELF_API_KEY = "…" }

Docker

Multi-arch, with SBOM and build provenance:

docker run -i --rm \
  -e AUDIOBOOKSHELF_URL=https://abs.example.com \
  -e AUDIOBOOKSHELF_API_KEY=… \
  ghcr.io/ni-c/audiobookshelf-mcp

-i is required — the protocol runs over stdin and stdout. There is no port to publish. More client recipes, including how to keep the key off the docker run command line, are in the client guide.

Through mcp-hub

A client that cannot spawn a local process — ChatGPT connectors, Claude on the web, Cursor, LibreChat — reaches audiobookshelf-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": {
    "audiobookshelf": {
      "command": "npx",
      "args": ["-y", "audiobookshelf-mcp"],
      "env": { "AUDIOBOOKSHELF_ALLOW_TOOLS": "essential" },
      "denyTools": ["delete_*"]
    }
  }
}

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

Every tool declares an outputSchema and answers with structuredContent alongside the text block, so a client can use the result without parsing prose. Seven tools that answered with a sentence — "Collection col_1 deleted." — now answer with the fields as well.

The tools that report library metadata carry untrusted: true and source: "audiobookshelf" as fields: book descriptions pulled from metadata providers, podcast feed summaries and episode titles are all written by someone else, and so are the bookmark titles and selected tags of an account and the titles of a library's longest and largest items. The rest are without it — an id this server was given, a position it was asked to store, the version string of the instance, and the library names and folder paths the operator typed.

The documents are described as open objects with the top-level keys this server builds. detail: "full" hands the API record back whole, so the same tool answers with far more keys than it names — and the SDK validates each result against its schema before it goes out, which is exactly why a strict shape would be wrong here.

get_personalized_shelves answers {items: [...]} rather than the bare array the API sends: a schema whose root is an array is served to a 2025-era client rewritten as {result: …}, so it would otherwise answer in two shapes.

Reading

Writing

👤 asks a person through MCP elicitation · falls back to a two-call confirm_token where the client cannot show a dialog.

Response size

Audiobookshelf returns very large objects — an expanded library item carries every audio file, track and chapter with full ffprobe metadata. Every tool that returns media therefore answers with a compact projection by default and accepts detail: "full" for the raw object.

Three bounds, because one is not enough:

• A response ceiling of 5 MB. content-length is checked before a byte is read and a chunked body is counted while reading, so an oversized answer is refused rather than parsed.
• A result ceiling of 100 000 bytes, applied in jsonResult — so it covers detail: "full" too. Whole entries are dropped, never characters: a truncated document is not a smaller answer, it is an unparseable one. The result then carries a truncated block naming what to call instead.
• A cap on embedded members. A compact collection or playlist embeds the first 25 of its books or entries and reports the real count; get_collection and get_playlist return the whole membership for one of them.

list_library_items pages properly, with limit and page. Seven listing tools have neither — list_libraries, list_authors, list_tags, list_genres, list_collections, list_playlists and list_bookmarks — because the Audiobookshelf routes behind them return everything in one answer and take no paging parameters. library_id narrows the two collection routes; the rest are bounded by the ceilings above.

Filtering

list_library_items takes filter_group plus filter_value and builds the base64-encoded filter parameter the API expects. The valid values come from get_library_filter_data. A valued group without a value is rejected, because Audiobookshelf would silently answer with the unfiltered library instead.

filter_group="authors",  filter_value="<author id>"
filter_group="progress", filter_value="finished" | "in-progress" | "not-started" | "not-finished"
filter_group="issues"    (standalone, no value)

Not exposed, on purpose

No playback. Pausing, seeking and playing are a session state machine that belongs in a real client. set_media_progress covers "mark this finished" and "jump me to chapter 12"; your phone does the playing.

No administration, even with an admin key. There is no tool for user management, server settings, backups, cache purging, filesystem browsing, library or item deletion, metadata rewriting or file uploads. Those endpoints exist in Audiobookshelf; they are simply not wired up here, because the blast radius of a confused or manipulated model is bounded by the tool list.

Safety

• Read-only mode. AUDIOBOOKSHELF_READ_ONLY=true does not register the write tools at all, rather than refusing them at call time.

• A person is asked, not just told. The eight tools that can take something out — the three deletes, delete_bookmark, remove_books_from_collection, remove_items_from_playlist, and update_collection / update_playlist when they are asked to reorder, which replaces an order nobody can reconstruct — raise a real dialog through MCP elicitation, which the model cannot answer on its behalf. A plain confirm: true flag could be set by the model on the first try, or be talked into it by text coming out of the library.

  Where the client cannot show a dialog they fall back to a single-use token bound to the exact targets and expiring after five minutes. That fallback proves the call was made twice with the same arguments and nothing more, and the text says so rather than implying somebody approved. ELICITATION=false takes it deliberately; it never removes the guard. See Asking a person.

• Confirmation prompts never quote API content. Collection and playlist names are user-supplied text and are read by a model, so the prompts name ids and counts only.

• A 200 that is not JSON is an error. Returning the body as a string made an SSO portal or a captive proxy in front of the instance look like an empty library rather than like a failure.

• Untrusted content is marked. Book descriptions come from metadata providers and podcast summaries come from RSS feeds — third parties write them. Every result carrying such content is labelled as data, not instructions.

• The API key is deleted from the environment once the configuration has been read, so it is not visible to child processes or in /proc/<pid>/environ.

• No redirects are followed (redirect: 'error'), so the Authorization header cannot be replayed against another host, and every request has a 15 second timeout.

• Ids are validated before they enter a URL path.

• Upstream error bodies are sanitized: HTML error pages are dropped, anything else is truncated to 2000 characters.

• Progress updates send whitelisted fields only. The Audiobookshelf endpoint applies its payload to the progress record wholesale.

• What this server cannot do, by design: no user management, no server settings, no backups, no cache purging, no filesystem browsing, no library or item deletion, no metadata rewriting, no file uploads.

One caveat that comes from Audiobookshelf itself: removing the last entry from a playlist deletes the playlist. remove_items_from_playlist says so in its result when it happens.

Documentation

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

Development

npm install
npm run lint          # oxlint + prettier --check
npm run build         # tsc
npm test              # vitest
npm run test:coverage # with thresholds
npm run docs:tools    # regenerate docs/reference/tools.md from the registered tools

The table above is hand-curated; the complete tool reference with every parameter is generated from the code, and CI fails if the committed copy is stale. See CONTRIBUTING.md for a throwaway Audiobookshelf you can safely write to — the write tools change progress and bookmarks on the API key's own user, so don't develop against a library you care about.

The tool definitions were derived from the Audiobookshelf server source (server/routers/ApiRouter.js and the controllers) rather than from api.audiobookshelf.org, which is out of date in several places — the filter data endpoint is /filterdata not /filter, progress updates are PATCH /api/me/progress/:id not POST /api/me/progress, and bookmarks live under /api/me/item/:id/bookmark.

Releasing

Tag-driven, no manual publish step:

1. Move the [Unreleased] entries into a new ## [x.y.z] - YYYY-MM-DD section in CHANGELOG.md and bump package.json.
2. npm run lint && npm run build && npm run test:coverage.
3. Commit, then a signed annotated tag: git tag -s vx.y.z -m "vx.y.z".
4. git push origin main vx.y.z.

release.yml then runs the tests, publishes to npm with provenance via Trusted Publishing (no token secret involved), creates the GitHub release from the CHANGELOG section, and publishes to the MCP registry as io.github.ni-c/audiobookshelf-mcp. ci.yml pushes the multi-arch image to GHCR on the same tag.

If the registry step fails, fix it on main and dispatch the Publish to MCP Registry workflow — do not re-run the tag job, which would check out the old tree.

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