abs-mcp

command module
v0.5.0 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Sep 27, 2026 License: GPL-3.0 Imports: 4 Imported by: 0

README

abs-mcp - an Audiobookshelf MCP server, CLI and Go SDK

GitHub release Go Version License build tests lint coverage

An MCP server, CLI and Go SDK that audits an Audiobookshelf library for the things that actually go wrong, and fixes what it finds - from Claude Code, Claude Desktop, or any other MCP client.

There are several Audiobookshelf MCP servers, and they do a useful thing: expose the API as tools, so a model can browse your library and read your progress. This one does that too, but the reason it exists is the layer above: 19 audits, each a sweep over the whole library for one specific thing that goes wrong in a real collection, returning a worklist rather than a dump, and naming the tool that fixes it.

This is not a demo. It has been battle-tested on a real collection: a large library, collected over years from every source and matched by hand or not at all, was cleaned up with these tools driven from Claude Code. Hundreds of titles were matched to the right store edition and checked against it, wrong matches caught by the folder the collector had named, author and narrator records merged and photographed, genres and tags brought to one vocabulary, and series names, numbering and titles brought to one style across the whole shelf in a sitting - every "The" dropped from a series label, every series past nine books zero-padded, every book linked to the series its folder names, every title that was really a series name replaced with the one the folder carried. Each pass was an audit, a review of the worklist, and a batch of edits. Most of the audits exist because that library had the problem.

What else is in the box
  • The whole API, as tools. 98 tools over all 202 Audiobookshelf routes, so everything an audit finds can be fixed from the same session: matching, covers, chapters, embedding, renaming a genre everywhere it is used, merging duplicate authors.
  • A Go SDK. lib/abs is a complete Audiobookshelf API client - 205 methods, no dependencies outside the standard library, no knowledge of MCP - useful on its own, whether or not you care about AI.
  • Tested against a real server. Every tool and every client method runs against an actual Audiobookshelf in Docker, and the suite fails if a registered tool has no test. Seven response-shape bugs in this client were found that way and could not have been found any other way, because Audiobookshelf publishes no OpenAPI spec and its public API docs say they are unmaintained.
The audits
audit what it catches
audit_all every audit in one call, counts only, so one call says where a library needs work - start here after a scan (deep adds the four that fetch something for every item, and counts audit_chapters and audit_whitespace whole)
audit_unmatched books never matched to a metadata provider: no asin and no isbn, so nothing else can be filled in automatically; a book tagged zz-provider:none, checked and found on no provider, is not reported
audit_missing items with one metadata field left empty - field is cover, description, narrator, series, author, genres, year, publisher, language or chapters; description also catches a stub, a credit line such as "Read by Paul Heck", or a bare url
audit_issues items whose folder is missing from disk or holds no playable media: broken records, not metadata gaps
audit_no_audio items with no audio tracks at all, usually an ebook-only folder that landed in an audiobook library
audit_path items whose folder name disagrees with their title or author once subtitles, series prefixes and edition markers are set aside: a wrong match, a chapter tag left as the title, a pen name, or a book filed under another author; with files the audio filenames are compared too
audit_chapters everything wrong with a book's chapters: chapters that start past the end of the audio (a file taken out of a book leaves them behind), chapters out of order, chapters that end well before the audio does (a track added), and one chapter over a long book, as unnavigable as none but invisible to audit_missing; item_chapters_set fit=true fits a book's own chapters back to its audio
audit_duplicates items that appear to be the same work: copies joined by any key they share - asin, isbn, or title+author, so a matched copy and an unmatched one of the same book are one group - but never two different asins. A title, or an isbn (the print edition's, which every recording of it carries), groups no copies anything tells apart, as a false group costs a deletion and a false split only a look: readers sharing no name (the narrator field, or a folder's brackets - (Weiner), [John Lee] - unless they are a year, an edition, a format, a source, a language, a note on the copy or the author), a number in their names (Vol. 17, CD1, Book One) or a place in one series (Honor - 01 - Above All, Honor - 02 - Honor Bound, both titled Honor), folders naming different books, one copy holding no audio, or more than 15% apart in length; a title over 50 copies share is a placeholder and joins none. Nothing is dropped: split lists every title and isbn whose copies were kept apart, with how many copies, the first twenty parts and why, so a part that proves one recording is still in sight. Beside the groups, candidates: pairs no key joins that are probably one recording - an edition label on the same audio (Ender's Game (20th Anniversary)), the same title filed under another author, one copy's album tag naming the other, split files beside a single file, or one title's copies kept apart only by their folders, naming no different readers, within 4% (Treason, A Planet Called Treason) - each to confirm with item_compare_audio. And incomplete: a copy sharing a title, an asin or an isbn, more than 15% shorter than another copy, holding only some of its tracks in play order - not a second copy but a bad one, reported apart from the groups with the first five files it lacks
audit_abridged books that are probably abridged but not marked so: the length of the store's abridged edition, far shorter than every unabridged one, or, for two readings of one book in the library, chapter lengths cut unevenly where two unabridged readings keep a steady ratio; a book under an hour long and under a fifth of the shortest unabridged edition, when the store sells no abridged edition under twice its length, comes back as shorter_work, probably a story sharing a collection's title and not abridged; productions (dramatised, full cast, a radio or audio drama, play or theatre, a Hörspiel, in the names or the reader credit, and not also saying unabridged) are set aside on both sides, and a copy the length of the store's production is that production (one store search per book, so it pages and audit_all runs it only with deep)
audit_series everything wrong with series: a book missing between the lowest and highest number the library has (interior gaps only), with the unlinked book on the shelf that fills it when there is one, and what is still missing once two spellings are read as one series; two series that are one series spelled two ways (The Chronicles of Amber vs Chronicles of Amber), each with its author, fixed by series_merge; a book whose folder says - 1 - while its series says #4, a book whose folder puts it in a series it is not linked to, or two titles at one number; names that are not series names, ending in the word Series, called what the author is called, carrying a book number, or with a stray ™ or a space before a colon; titles that are the series name with a number, or carry it beside the real title, where the folder says what the title is; one series' folders written two ways on disk (Pandora's Star beside Commonwealth Saga - 02 - Judas Unchained, a double space, - 3 - beside - 03 -), each with the folder name that would match; and with articles=true the names that open with The, A or An
audit_spelling the same value spelled several ways across genres, tags, languages and publishers: Sci-Fi vs sci fi, en vs eng vs English, HarperAudio vs Harper Audio, and spellings a letter apart such as Romance vs Romances
audit_authors everything wrong with authors: a book whose author field holds its title; records never matched, with no photo, with no books, or whose biography opens with someone else's name (a wrong match: Sarah Diemer carrying Sarah Miller began writing...); and records that are one name spelled two ways (C Z Dunn vs Christian Dunn)
audit_narrators everything wrong with the narrator field: names that wrote some books and read others (one volume written and the rest of the series read is author and narrator swapped on import; one written and many read is a narrator credited as co-author), and the spelling checks on narrators - Read by Jim Dale, Narrator....Jim Dale, Fajer Al vs Fajer Al-Kaisi, two names in one value, a stray Ph.D.
audit_genres genres and tags against one rule, genres broad and few, tags fine and many: placeholders like Audiobook in either field, a category path written as one value (Science Fiction & Fantasy, Fantasy) with its parts, genres on too few books to be a genre, tags that repeat the book's genre, books with no genre; each with the metadata_rename call that fixes it
audit_whitespace spaces where a name should not have them - a double space, a space at either end or before the extension (End Credits .m4b), a space before a colon or the look-alike ꞉, a tab or non-breaking space (the ideographic space Japanese titles use is left alone) - in titles, subtitles, author, narrator and series names (read from the library's own lists, so Jane Doe, Ph.D. stays whole), every folder of the path, disc folders too, and every file in a book's folder. Each row shows the spaces and the name put right, with the item's title beside a folder or file (a double space often marks a dropped colon), the record id to fix an author or series by, and, on a folder or file row, under taken any suggestion already used beside it, which a plain rename would overwrite
audit_unembedded books whose audio files do not carry the library's metadata in their tags, never embedded or stale since the last edit: what item_embed_metadata is due for
audit_covers everything wrong with cover art: missing, not square, or too small (reads every cover's header, so audit_all runs it only with deep); with banner=true, the "Only from Audible" ribbon across the bottom-right corner, found by colour and angle; and with store=true, compared with the store the book was matched to by perceptual hash: the same picture bigger is an upgrade with the full-size url, another picture is differs with both images to look at, a window of matched books per call, paged by offset
audit_matched matched books whose asin is not the recording on disk: the provider's record for the asin compared on title, duration and narrator, so a match applied by title alone shows up as duration_off or narrator_differs; with fields every field is compared and the differences listed with both values, which is what override_details would change (one provider request per book, so it pages and audit_all runs it only with deep)
audit_podcast_stale_feed podcasts whose newest episode is over 90 days old, or whose feed was never checked: the show ended, or the feed url is dead
audit_podcast_no_episodes podcasts with nothing downloaded

The design principle: detection is code, correction is judgment. The server runs cheap deterministic checks over the whole library and produces worklists; the AI reasons only about the anomalies. Every response is a trimmed projection of what a decision needs, never the raw API object (an expanded library item carries every audio file, track and chapter with full ffprobe output).

Installation

go install github.com/katbyte/abs-mcp@latest

Requires Audiobookshelf 2.26 or newer (earlier versions have no API keys).

Configuration

All options can be passed as command-line flags, environment variables, or via a configuration file.

Variable Flag Description
ABS_SERVER --server, -s Audiobookshelf URL, e.g. http://nas:13378
ABS_TOKEN --token, -t API key (Settings → Users → API Keys)
ABS_READ_ONLY --read-only register only tools that never change server state
ABS_ENABLE_DELETE --enable-delete register the tools that delete items, episodes, authors, collections and playlists
ABS_PROVIDERS --providers metadata providers to ask in order when a call names none, the store the books were bought from first: audible.ca,audible; default the library's own provider
ABS_PROVIDER_TAG --provider-tag prefix of the tag that records which store a match came from, zz-provider: by default so it sorts last in the tag list; off writes and reads none
ABS_TOOLSETS --toolsets groups of tools to register, default core: all, core, curation, listening, podcasts, organise, admin, or a resource family like item (core is always included)
ABS_ALLOW_TOOLS --allow-tools only register these tools (names, library_* globs, or essential)
ABS_DENY_TOOLS --deny-tools never register these tools (names or globs such as *_delete)
ABS_LOG log level (WARN default; DEBUG, TRACE, ...)
ABS_LISTEN --listen serve MCP over HTTP on this address (e.g. :8080) instead of stdio
ABS_AUTH_TOKEN --auth-token bearer token required on the HTTP endpoint (required with --listen)
ABS_ALLOW_NO_AUTH --allow-no-auth serve HTTP with no bearer token at all: anyone who can reach the port can use every tool

An API key acts as exactly one Audiobookshelf user and inherits that user's permissions: a key for a normal account cannot see libraries that account cannot see, and cannot scan, match or delete. Most write tools need an admin account; server_info reports what the key can do.

Configuration File

You can place a .abs-mcp file in your home directory ~/.abs-mcp (for global settings) and in your current directory ./.abs-mcp (for per-project settings). Both are read, the project's over the home one, so a project file only needs the settings it changes. Keys are the environment variable names, with or without the ABS_ prefix, in env format:

SERVER=http://nas:13378
TOKEN=ey...
READ_ONLY=true

A flag or an environment variable beats either file.

Usage

Quick connectivity check:

abs-mcp info
Register with Claude Code

.mcp.json:

{
  "mcpServers": {
    "audiobookshelf": {
      "command": "abs-mcp",
      "args": ["serve"],
      "env": {
        "ABS_SERVER": "http://nas:13378",
        "ABS_TOKEN": "..."
      }
    }
  }
}

Or from the shell:

claude mcp add audiobookshelf -e ABS_SERVER=http://nas:13378 -e ABS_TOKEN=... -- abs-mcp serve
Run as a service (HTTP transport)

serve --listen :8080 serves the MCP Streamable HTTP transport at /mcp (plus GET /healthz) instead of stdio. ABS_AUTH_TOKEN is required: clients must send Authorization: Bearer <token>, and the server refuses to start without one unless ABS_ALLOW_NO_AUTH=true says that anyone who can reach the port may use every tool. Register it from any machine:

claude mcp add --transport http audiobookshelf http://nas:8080/mcp \
  --header "Authorization: Bearer $ABS_AUTH_TOKEN"
Docker

Releases publish a multi-arch (amd64, arm64) image to ghcr.io/katbyte/abs-mcp, tagged vX.Y.Z, vX.Y and latest. docker-compose.yml is the default always-on deployment: it runs that image and reads secrets from a gitignored .env (copy .env.example). Adjust ABS_SERVER and TZ in the compose file, then:

cp .env.example .env      # fill in ABS_TOKEN and ABS_AUTH_TOKEN
docker compose up -d

make docker builds the same image from source, tagged abs-mcp, with version info from git. The image is alpine-based (so docker exec -it abs-mcp sh works), runs as a non-root user and has a healthcheck against /healthz. The binary is the entrypoint, so docker run --rm ghcr.io/katbyte/abs-mcp info works as a connectivity check with the ABS_* variables passed via -e.

MCP Tools

Tools are named resource-first (library_*, item_*, user_*...) so they group by what they act on. Every tool carries MCP annotations (read-only or destructive) and tools that change server state say so in their descriptions. Wherever a tool takes a library, item, author, series, collection, playlist or user it accepts a name as well as an id; an ambiguous title comes back as an error listing the candidates.

Resource Tools
server server_info (connectivity, permissions, libraries, providers and server-wide totals), server_tasks, server_sessions, server_backups, server_backup_create, server_tags
libraries library_list, library_get (in depth, with statistics), library_create, library_edit, library_search, library_items (the server's own filters and sorts: genre, tag, author, series, narrator, progress, tracks...), library_recent, library_filters, library_scan
audits the 19 audits in the table above: audit_all, audit_unmatched, audit_missing, audit_issues, audit_no_audio, audit_path, audit_chapters, audit_duplicates, audit_abridged, audit_series, audit_spelling, audit_authors, audit_narrators, audit_genres, audit_whitespace, audit_unembedded, audit_covers, audit_matched, audit_podcast_stale_feed, audit_podcast_no_episodes
items item_get (with optional chapters and files), item_edit, item_batch_edit (same fields across many books; add_tags and remove_tags edit each book's own list), item_rescan, item_embed_metadata, item_compare_audio (are two items the same recording? It samples five points of one book and finds each in the other by its loudness, allowing for a copy a few percent faster or slower: a re-encode, a split copy or another edition's label scores high, another narrator low. It reads about half of the second book over the network and needs ffmpeg where abs-mcp runs; the Docker image has it)
matching item_match (candidates from a provider), item_match_apply, item_match_batch → item_match_apply_batch (a window of books scored against the provider, paged by offset, then the accepted rows applied by asin; both apply tools take override_details with a keep list of fields to put back afterwards, or smart, which fills the empty fields and then decides each remaining difference by rule, writing file-tag titles and company narrators over, keeping curated series and plain years, and reporting the rest for review; preview shows the decisions without changing anything. Every applied match records its store as a zz-provider: tag, which the audits and the batch search ask first, and item_match_tag backfills it for books matched before the tag existed), item_cover_search, item_cover_edit (url, file, or remove), item_cover_upgrade (the store's full-size cover when it is bigger and the same picture), item_chapters_set (explicit list, from Audible by asin, or fit to fit the book's own chapters to its audio)
authors author_list, author_get, author_edit (rename to merge duplicates), author_match → author_match_apply (look up on Audible, check the candidate, then apply by asin), author_image_set
series series_list, series_get, series_edit, series_merge (one series into another, numbers and other series kept)
narrators narrator_list
metadata metadata_rename (a tag, genre, narrator, author, language or publisher, everywhere it is used; renaming onto an existing value merges, remove drops it once confirm follows the preview)
collections collection_list, collection_get, collection_create, collection_edit, collection_books_edit (add or remove), collection_delete
playlists playlist_list, playlist_get, playlist_create (also from a collection), playlist_edit, playlist_entries_edit (add or remove), playlist_delete
podcasts podcast_episodes (one show, or the newest across the library), podcast_episode_get, podcast_episode_edit, podcast_check_new, podcast_feed_episodes, podcast_episode_download, podcast_downloads, podcast_search, podcast_add, podcast_settings
users user_get, user_in_progress, user_progress_get, user_progress_set, user_progress_remove, user_bookmarks, user_bookmark_edit (add or remove), user_history, user_stats (all-time or year in review), user_list (admin)

item_delete, podcast_episode_delete, author_delete, library_issues_remove, collection_delete and playlist_delete are only registered when --enable-delete / ABS_ENABLE_DELETE is set; the ones that erase files or many records at once (item_delete, podcast_episode_delete, library_issues_remove, and metadata_rename with remove) say what they would remove and change nothing until called again with confirm. --read-only registers the 58 read tools and nothing else, so a write tool is absent from tools/list rather than refused when called.

Choosing which tools load

The default is core: five read-only tools, about 1,200 tokens. The whole surface is around 25,000 tokens of tool definitions before anyone asks a question, which is a poor way to spend a client's context by default. --toolsets / ABS_TOOLSETS loads the groups a session actually needs, and core comes along with whatever else is asked for, because nothing else can find a library or open an item.

Curating a library needs ABS_TOOLSETS=curation - the audits and everything that fixes what they find. ABS_TOOLSETS=all restores every tool.

toolset tools with core ~tokens
core (default) 5 5 1,200
admin 13 18 2,900
organise 12 17 2,600
podcasts 11 16 3,200
listening 9 14 2,800
curation 48 53 17,800
all 98 98 24,600

Tokens are what the model sees: each tool's name, description and input schema, measured over a real tools/list at four bytes a token. Every tool also carries an output schema, another 34,000 tokens across all, but clients keep that to themselves to validate results rather than sending it to the model.

--toolsets also takes a resource family - item, podcast, library, user, audit, author, series, narrator, collection, playlist, server - which is every tool with that prefix:

ABS_TOOLSETS=all                # every tool, which was the default before 0.2.0
ABS_TOOLSETS=curation           # audits plus everything that fixes what they find
ABS_TOOLSETS=listening,podcasts # a client that plays things rather than curates them
ABS_TOOLSETS=audit              # read-only detection, nothing that writes
ABS_TOOLSETS=core,item,series   # core plus two whole families

abs-mcp tools prints what the current flags would register, grouped by toolset, and needs no server:

abs-mcp tools                   # the default set
abs-mcp tools --toolsets all    # every tool
abs-mcp tools --read-only -q    # names only
Narrowing further

--allow-tools and --deny-tools narrow whatever the toolsets left, and take comma-separated tool names, globs with a leading or trailing *, or the essential preset (library_list, library_search, library_items, item_get, user_in_progress, user_progress_get, user_progress_set):

ABS_ALLOW_TOOLS=essential
ABS_ALLOW_TOOLS=library_*,item_get,user_*
ABS_DENY_TOOLS=*_delete,server_*

A pattern that matches no tool aborts startup and names it, so a typo cannot silently hide a tool.

A typical curation session
  1. audit_all says where the library needs work; audit_unmatched lists the books never matched to a provider.
  2. For each, item_match returns candidates with duration, narrator and series; compare them with the item and item_match_apply candidate=N.
  3. audit_missing field=cover and item_cover_search / item_cover_edit fill the gaps.
  4. audit_missing field=chapters finds long books with no chapters; item_chapters_set pulls them from Audible by asin.
  5. audit_duplicates and audit_series show what to prune and what is missing.

Using the client on its own

lib/abs is a plain Go client for the Audiobookshelf API with no dependencies outside the standard library, and no knowledge of MCP. If you only want to talk to Audiobookshelf from Go, take it and ignore the rest:

import "github.com/katbyte/abs-mcp/lib/abs"

client, err := abs.New("http://nas:13378", os.Getenv("ABS_TOKEN"))
items, err := client.Items(ctx, libraryID, abs.ItemsOptions{Limit: 50})

It has 206 methods covering every one of Audiobookshelf's 202 API routes - libraries, items, authors, series, narrators, collections, playlists, progress, bookmarks, podcasts, provider search, RSS feeds, tags, genres, tasks, backups, playback sessions, notifications, email, API keys, sharing, settings and user administration. File downloads stream rather than buffer, so a multi-gigabyte audiobook does not have to fit in memory.

make apicheck reads the route table out of the Audiobookshelf source and fails if anything is missing, so the coverage claim is checked rather than asserted. Audiobookshelf publishes no OpenAPI spec and its public API docs say they are unmaintained, so the types here are written against the server source (see docs/README.md) and then proved against a running server - which is the only thing that catches the server changing shape underneath you.

Development

make            # fmt + build
make check-all  # build + unit tests + both live suites (needs docker) + every linter
Tests

make test is hermetic and fast. It covers the pure logic - filter encoding, formatting, gap arithmetic, the audit heuristics, tool registration - and two things that need a server but not a real one: every lib/abs request shape and response decoding against a canned server (lib/abs/requests_test.go), and the tools end to end over an in-memory MCP session against a canned Audiobookshelf (tools/handlers_test.go). The second is where the cases the live fixtures cannot reach live: a library with covers, inconsistent spellings, tagged and untagged audio files, more findings than the limit.

Everything else runs against a real Audiobookshelf in Docker, because a stub can only confirm what you already believed. Two suites, each in its own container:

Covers Command
integration/ the lib/abs client: that every response decodes with its fields populated make testacc-integration
acceptance/ the tools: name resolution, projections, audits, provider flows, journeys, and a smoke test of the built binary make testacc-acceptance
make testacc        # both, each in a throwaway container, torn down after
make check-all      # build + unit + both live suites + every linter
make cover          # all three suites, merged into one coverage number

The journeys (acceptance/journey_*_test.go) chain the tools the way a session does and read the server back after every write, because a 200 from Audiobookshelf is not proof: edits made while a scan runs; every fixable audit fixed, audited again and put back; a non-admin's playback reaching every listening tool; audit_all equal to each audit; every read-only lookup leaving the server byte-for-byte unchanged; accounts limited to one library or one tag; a library's whole life; writes done twice; every record got by id and by name; a podcast's whole life over a feed the suite serves itself through the provider proxy; metadata embedded, the record thrown away and the book scanned back from its file alone; a book deleted out from under collections, playlists, progress, bookmarks and someone's history; one turn's calls made at once on the same records; an account without rights calling every tool that changes the server; a match decided field by field, one asin on two books, and every page of every list read exactly once; books deleted with their files, a folder renamed under its book, a long book chaptered, and issues removed one library at a time; and an explicit book, a tag-limited listener's queue and a show holding one episode twice. They found some thirty tool bugs the per-tool tests had not, and the server behaviour behind them is listed in docs/README.md.

Those journeys, like everything else, drive tools.RegisterAll in process, which is every line of tool code the binary runs. What they never touch is the thin layer around it, where a breakage is silent: MCP over stdio uses stdout for the protocol, so one stray print or log line corrupts the stream and a client simply fails to connect with the whole suite still green. So acceptance/binary_test.go builds the real binary and speaks to it the way a client does: serve over stdio at the most verbose log level, with every line it writes to stdout checked for being a protocol message and nothing else; the flags, the environment and a .abs-mcp in the working directory each changing what a client lists; --listen serving /mcp behind the bearer check with /healthz open beside it, and shutting down on SIGTERM; and a start that cannot work - no server, no token, --listen with no auth token, a toolset that does not exist - failing with a message that names the problem rather than hanging.

Coverage has to span all three or it lies: go test -cover ./... reports about 40% for tools/, because almost everything real happens in the live suites behind the integration tag. make cover runs each into its own binary coverage directory and merges them with go tool covdata - stdlib tooling, no third-party merger - which is what the badge reports.

All 98 tools and all 205 client methods are exercised, 200 of them asserting a result rather than only that the call reached the server. The five that do not - sending an ebook by email, firing a notification, closing a device session, unlinking OpenID, syncing an offline session - need infrastructure a throwaway container has not got, and say so where they are written. Tool coverage is enforced rather than claimed: the acceptance suite records every tool it calls and fails if the server registered one nothing called, so a new tool cannot ship untested. Calls out to Audible, Audnexus and iTunes go through a record/replay proxy (lib/providerproxy), so neither suite needs a network:

make record         # re-record every cassette against the real providers
make record-check   # check the cassettes still match, without rewriting them

record-check compares the shape of live responses against the recordings - renamed fields, vanished fields, changed types - and ignores values, so it goes red when a provider changes its contract rather than when a chart position moves.

Fixtures are generated, never committed: scripts/abs-testenv.sh writes one-second silent files with ffmpeg under ~/.cache/abs-mcp (ABS_TEST_DATA to move them - not $TMPDIR, which Docker Desktop does not share), creates the libraries through library_create, fills them with library_scan and sets the metadata with item_edit - so building the fixtures is itself part of the coverage. Three libraries are clean, so the audits have something to leave alone; the fourth, Messy, is seeded with every defect they exist to find - a series spelled two ways, a gap whose missing book sits unlinked in a series folder, a narrator's name misspelt on half the books, a folder naming another book, a duplicate, a cover wearing the ribbon - and each audit has a test against it. scripts/abs-testenv.sh fixtures writes just the audio tree if you want to look at the layout. Requires docker, ffmpeg and jq; the suites skip when ABS_SERVER and ABS_TOKEN are unset, so they never fail for want of a daemon.

The Audiobookshelf API reference is the server source, not the public docs; see docs/README.md.

Documentation

Overview

Package main implements abs-mcp, an MCP server and CLI for curating an Audiobookshelf library.

Directories

Path Synopsis
Package cli implements the abs-mcp command line interface: the cobra commands, flag and config handling, and the MCP server exposing Audiobookshelf library tools to AI clients.
Package cli implements the abs-mcp command line interface: the cobra commands, flag and config handling, and the MCP server exposing Audiobookshelf library tools to AI clients.
lib
abs
Package abs is a minimal client for the Audiobookshelf HTTP API.
Package abs is a minimal client for the Audiobookshelf HTTP API.
audiosample
Package audiosample reads stretches of a book's audio from an Audiobookshelf server: seconds or minutes from anywhere in the book, as mono PCM at the rate asked for, across its files in play order.
Package audiosample reads stretches of a book's audio from an Audiobookshelf server: seconds or minutes from anywhere in the book, as mono PCM at the rate asked for, across its files in play order.
providerproxy
Package providerproxy is a record/replay HTTP proxy for the metadata providers Audiobookshelf calls out to.
Package providerproxy is a record/replay HTTP proxy for the metadata providers Audiobookshelf calls out to.
Package tools defines the MCP tools exposed by abs-mcp, split by the resource they act on (server.go, libraries.go, items.go, ...).
Package tools defines the MCP tools exposed by abs-mcp, split by the resource they act on (server.go, libraries.go, items.go, ...).

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL