Where the data comes from

Four sources are asked at once and merged into one entity:

Source	Brings
TMDB	overviews, stills and translations, in the configured language
TheTVDB v4	the season and episode numbering, absolute numbers, air-order hints, broadcast time of day, per-country certifications
Skyhook (Sonarr) / api.radarr.video	the same view the arr stack would have got on its own, including ratings it carries and TMDB does not
Fanart.tv	transparent logos, clearart and banners — artwork only

Only TMDB is really needed to start. The rest are optional and on by default; each costs one extra call per refresh.

The merge is field by field, in the order of the setting sources.priority — put in order on the Sources page, and given its first value by AMS_PROVIDER_PRIORITY:

Scalars take the first provider that has anything, so a lower-priority source fills a gap rather than overwriting an answer.
Artwork, alternative titles, ratings and translations are unioned.
Credits come from one provider; nothing identifies a person across providers, so a union would list the same actor twice.
Episodes and seasons are the exception, and the one rule worth knowing.

Whose episode numbering wins

The list of episodes and seasons always comes from TheTVDB, or from Skyhook which republishes its numbering. Every other provider may fill fields on those episodes; none may add one.

This is not configurable, because it is not really a preference. A client addresses a series by its TVDB id and expects the numbering that goes with it, and the providers do not agree on where seasons end — so serving TMDB’s boundaries under a TVDB id makes Sonarr map files to the wrong episodes.

Taking the union of both is worse still: episodes that are the same hour of television sit at different numbers in each scheme, so none of them match and every one is kept twice. That is not hypothetical — it is what this server did before the rule existed, and for One Piece it answered with 2352 episodes for a show that has 1179.

Measured against the live APIs with all four providers configured:

Series	Numbered episodes	With absolute numbers	With overviews
Breaking Bad	62	62	62
Attack on Titan	89	89	89
One Piece	1179	1178	1178

Removing TMDB and leaving the other three changes none of those counts — that is the point of the rule — but costs the artwork TMDB carries: Breaking Bad drops from 82 images to 69.

A real Sonarr 4.0.20 served by this server stored those same counts with no gap in the absolute numbering. A real Radarr 6.4.4 searched by id and by text and added films with their studio, genres, certification and IMDb/TMDB/Trakt ratings. Jellyseerr’s ten TMDB endpoints all answer through the relay.

Enrichment is asking a provider on every fetch and merging what it says. It is on by default and costs one extra call per refresh per provider. Fallback is asking only when nothing else could answer, so with enrichment on it never fires — the answer is already in hand, and a second call would buy it twice.

Both are settings, changed under Settings without a restart; the AMS_* variables only seed them the first time the server starts.

A search is a ladder rather than a fan-out, because search results are cheap to get and expensive to merge: what is already stored, then TMDB, then TheTVDB, then Skyhook — each rung tried only while the ones above it found nothing. So a series TMDB has never heard of is still found, and a deployment with no TMDB key at all still answers.

If you redirect skyhook.sonarr.tv or api.radarr.video at this server through your resolver rather than per container, it resolves those names to itself and would call itself forever. It detects that and answers 508 with the fix in the message — but point the upstreams elsewhere, or turn enrichment off, when that is your setup.

Who gave what, and asking one source again

Every refresh records, field by field, which provider’s value was kept, which others said the same and which said otherwise; the pictures are counted by source, and the episode list names whoever numbered it. The work editor shows it beside each field: the provider’s name, Multi-sources with the count when several agree (Skyhook is not counted beside the TheTVDB it republishes), Locked for an edit by hand, Kept for a value no source gives any more, No source when nobody gave one. None of this reaches a visitor. Administration › Sources lays out the rules themselves: who can supply what, in the configured order, and who replaces the rest.

The editor’s Sources panel asks chosen providers again, and only them. The others are handed back what they gave last time, so the ordinary merge weighs the fresh answers against them by the ordinary priority: a sync from TheTVDB alone takes its new runtime without handing it a title TMDB outranks it on. Locked fields do not move, the refresh schedule and the work’s identifiers stand, and the episode list stays whoever numbered it unless that provider gives one anew. Some sources come along with the ones asked: TVmaze with the provider that numbers the episodes, whose broadcast instants it corrects, and that provider with any source that fills the episodes in, so the list is merged as a refresh merges it. A sync that finds the work written by something else meanwhile — a refresh, another sync — writes nothing and says so. A work refreshed before this was recorded needs one full refresh first. The same through the API: GET /api/v1/items/{id}/provenance, POST /api/v1/items/{id}/sync with {"sources": ["tvdb", "fanart"]}, and GET /api/v1/sources/rules.

Further sources

Six more, each off until switched on under Settings → Further sources. None needs a key.

Source	Brings	Costs
TVmaze	the moment each episode aired	two calls per series fetch
AniList	its score, the romaji, native and English titles and synonyms, the main studio, an adult flag — for anime	one call per anime fetch
MyAnimeList	its score and titles — for anime; through its own API when `AMS_MAL_CLIENT_ID` is set, through Jikan otherwise, and through Tenrai, a mirror written to Jikan’s schema, when the one asked does not answer (`mal.fallback`, on; `AMS_TENRAI_UPSTREAM`)	one call per anime fetch
IMDb	IMDb’s rating, for every work with an IMDb id	one download a day
Fankai	the Fan-Kai productions — anime recut into films, each as a series with its sagas as seasons — from Fankai’s own metadata service	one call per saga, and three more per production fetch
Wiki Fankai	for each Fan-Kai, the anime it was cut from and the Fan-Kai that follows it — only while Fankai is on	one call per Fan-Kai fetch

TVmaze’s servers take fifteen to twenty seconds, now and then, over what their cache does not hold — the first lookup of a series — whatever the pace of the calls. A refresh or a seed gives TVmaze a minute and asks once more after a pause; a client waiting on this server gives it twenty seconds, and the fetch goes without it. Its API keys are not asked for: they open its user API, a premium member’s own data, and change nothing of the public one’s limit, which is per address.

None of the first four changes the shape of an answer: Sonarr and Radarr are served the same fields as before, some of them now more accurate. Fankai adds works instead — ones no other source has.

TVmaze corrects airDateUtc, the moment after which Sonarr counts an episode as aired. TheTVDB keeps one broadcast time per series and Skyhook stamps it on every episode, so a show that changed slot is wrong for everything before the change. TVmaze keeps a time per episode. Measured against Skyhook: all 62 of Breaking Bad’s episodes are placed at 21:00 ET, where 54 aired at 22:00 — the last eight did air at 21:00, and on those it is TVmaze that is an hour late — and 35 of The Big Bang Theory’s move, season 3 having aired on Mondays at 21:30 rather than Thursdays at 20:00. Its time is taken only where its broadcast date matches TheTVDB’s for the same season and episode. It never adds, removes or renumbers an episode.

AniList and MyAnimeList are asked about one entry: the one that stands for the whole work, as the Fribb anime-lists id map places it — the first season’s, for a series both sites split by cour. The map is downloaded weekly while either source is on, and it also answers Sonarr’s mal: and anilist: searches, which is how its AniList and MyAnimeList import lists find a series; those used to find only series already stored here. Neither site is ever used for numbering. Sonarr ignores alternative titles, so the romaji ones serve Radarr, which matches anime film releases against them, and this interface. For anime, their main studio replaces TMDB’s, which is only the first production company TMDB lists.

IMDb’s rating is what Sonarr is given for a series — Skyhook’s rating is IMDb’s, republished — and otherwise the rating with the most votes behind it. The daily list keeps that figure current and supplies it when Skyhook and Radarr’s service are off. Only the ratings of works stored here are kept, and the list is fetched again within the hour when works are added.

Fankai lists the Fan-Kai productions: anime recut into films by the Fankai team, which neither TheTVDB nor TMDB carries. Each production is a series, its sagas are its seasons, and its films are its episodes, numbered as Fankai names its files — Horimiya Kaï.S01E02.MULTI.1080p.x265-FANKAI.mkv is season 1, episode 2, and a second saga carries on from the first: S02E08 is the eighth film. A production turns up in Sonarr’s search under its own name — Naruto Shippuden Yabai, Black Lagoon Henshū — or as fankai: and its id, and on the import page; nothing is fetched until it is asked for, so Fankai’s catalogue never lands here whole. Sonarr is handed an id of this server’s own, Fankai’s plus 200 000 000, the way a TMDB-only series gets one; Plex, Jellyfin and Kodi get the production through the NFO export, with Fankai’s id as a uniqueid and its theme music beside it as theme.mp3, which a work’s page also plays on demand. The kaïeur and the voice cast are its credits; the kind of recut — Kaï, Yabai, Henshū — is a tag rather than a genre. One call a second, and the listing is kept and revalidated by its ETag.

Fankai’s id is its metadata service’s. The website numbers its productions otherwise — Horimiya Kaï is series 33 of the one and production 101 of the other, where 33 is Black Lagoon — behind a sign-in, and nothing public says which page is which. So the id is shown without a link, and a production’s homepage is its page on the Fankai wiki below, when that source is on.

The Fankai wiki says what each Fan-Kai was cut from. Fankai’s service leaves a production’s ids blank; the community’s wiki, at fan-kai.fandom.com, links each one’s AniList and MyAnimeList entries and names the Fan-Kai that follows it. With it on, a Fan-Kai’s page leads to its anime — to the anime’s own page once Sonarr holds it — and to the next Fan-Kai, and the anime’s page leads back to every Fan-Kai cut from it. The wiki keeps a page per cut, so a production is matched on its name and on its kaïeur, never on a guess between two. With AniList on too, the anime’s title, year and poster come from there. The ids are never given to the Fan-Kai itself: sharing one would merge it into the anime.

The lists and their last download are shown under the switches, with a button to download one now rather than wait for its turn, and every download is a job. Their terms: TVmaze’s data is CC BY-SA; IMDb’s datasets are for personal, non-commercial use; AniList’s API is free for non-commercial use and asks not to be crawled or stored wholesale — this server asks only about works a client requested; Fankai publishes its metadata for its own productions, and is asked about one at a time; the Fankai wiki’s content is CC BY-SA. Every public page credits the sources that are on, with the notices TMDB and IMDb require.