Skip to content

Latest commit

 

History

History
1124 lines (1004 loc) · 93.9 KB

DATA_POLICY.md

File metadata and controls

1124 lines (1004 loc) · 93.9 KB

Data policy

Where Everloot's data comes from, what we refuse to use, and why. It applies to the addon, the site, the MCP server and every contribution. If a pull request adds data, it follows this page.

This is our reading of the rules, not legal advice.

2026-11-10 owner decision: wiki names and partial item stats. An NPC with one or two distinct installs reporting a unanimous name may show that name when it exactly matches a Warcraft Wiki NPC page's infobox name (or page title) after the item name normalization: Unicode NFC, case, spacing and apostrophe style. The 2023-11-13 archive's npcbox pages supply a normalized name set, with page paths, in the private build-stamped wiki-npc-names.json catalog (everloot.wiki-npc-names/2). The website compares the reports it already loads, honoring the existing recheck cutoffs. Disagreeing reports, no reports or no wiki match keep "NPC N". Three agreeing installs take precedence as before. The name carries the same Name from the Warcraft Wiki label as items and links the source page; the same name appears in headings, breadcrumbs, metadata, JSON-LD, social cards and site hover cards, and counts toward sitemap eligibility. Only counts are public, never reporter identities. This is a website rule; the API, MCP, bulk exports and embeddable cards retain their existing agreement rules. Wiki names are CC BY-SA 4.0. The legacy id-keyed spell wiki path remains inert: the current dump has no usable spell ids. Spell names from the current game files and player agreement work as before.

Item stat values reported by one or two distinct installs are shown on the item page, hover card and /api/tooltip, labelled unconfirmed, seen by N player(s), with an invitation to install the addon at /addon to help confirm them. Zero reports keeps [?]. Partial reports that disagree keep [?]. Three agreeing installs still define confirmed. Build selection and latest vote per install follow the existing stat rules; an older confirmed or contested build is not replaced by a newer partial build. Only counts are public, never install ids or who reported a value. The trusted-install / Everloot's own install exception does not confirm a stat: one own-install report is still unconfirmed. Other publication thresholds remain as before. This decision supersedes earlier wording that all partial stat values stay hidden.

What we publish

What players see in the game. The Everloot addon records what happens around you while you play: drops, vendors, auction prices you browse, recipes, trainers, gold in and out. The list of your installed addons is recorded only if you turn it on with /everloot addons on, and it is sent only from the website's upload page, after the preview names each addon. You choose when to upload. On the website you preview everything before it is sent. The Desktop App sends anonymous open-world events, and census rows if you allowed the census on its setup screen; the command-line uploader sends only the anonymous events. Neither sends named data (guild, gear, gold, collection) or the installed addon list: those stay on your machine. Player rows are a third case: the Desktop App sends them only while "Send the player pages with the uploads" is on in its settings, and while it is off they stay on your machine like the named data. From 17 November 2026 that setting starts on in a new Desktop App install; an install that already exists keeps its choice. Boss pull summaries (below) are a fourth: the Desktop App sends them, and the command-line uploader does not.

Recipe observations can include the profession API's recipe identifier and whether that recipe is learned, when the client supplies those values. They remain anonymous profession observations, not a linked character's recipe collection. Explicitly unlearned recipes do not establish craftable skill ranges.

What a Polestar Recorder file holds. A player who records with the Polestar Recorder addon (by fitzforge, GPL-3.0) can upload its Polestar_Recorder.lua from the website's upload page. Everloot reads the file's format with its own parser and uses none of Polestar's code. The file is converted in the browser, one session at a time, into Everloot's own kinds: where the player stood while an NPC's window was open (a quest giver, a quest turn-in, a gossip window offering quests, a vendor, a trainer, a flight master), stored as positionKind: "interaction" like the addon's; NPCs the player killed or saw as a rare on its nameplate, target or mouseover, where the player stood (positionKind: "player", never a pin); rares the minimap showed, at the rare's own position from the game's vignette API (positionKind: "vignette"); the dungeons and raids the player zoned into, with the instance's name and type and the last position outside it (instance_entrance); the boss encounters that ended inside, with their result and the bosses' ids (encounter); treasures the minimap showed, at the object's own position (object_sighting); the quests the player accepted with their title, level, giver, turn-in, reward items and the character's faction; and loot windows with items. Rares and kills keep the level and rank the game showed. The file holds no NPC level at a window and no reaction, so those fields are left out, never guessed. It does not say whether the game's loot API named a window's source either, so its loot windows stay out of the public drop rates: they are counted instead, and the NPC, object and item pages say how many were reported and are not yet in the rates. Everything else in the file stays on your computer: character names and realms, session names, notes typed with /psrec mark, Polestar's guide step titles, quest objective text, routes, flight paths, talents and the rest. A character becomes a hash in the browser, keyed with a random key that stays in that browser, so the name cannot be worked back from it; the server replaces that hash with an install id of its own, keyed with a secret: one half from the character, one from the address the upload came from. Recorder install ids that share either half count as one reporter, so a Recorder fact needs three characters uploading from three addresses, and one player's alts or connections never confirm each other. The upload keeps a keyed hash of the address, as auction scans do, and it is never published. Corroborated, these are Everloot's own data under CC BY-SA 4.0, like any player upload.

Uploaded observations are public as individual, anonymous records, with no character names and without the addon's install id. Each record carries a per-install pseudonym instead (pub- and 32 hex characters, a keyed hash of the install id that cannot be turned back into it), so anyone can count installs and check our counts. The facts the site states itself, such as drop rates, spawn points and recipe ranges, appear only once at least three installs have reported them. Addon lists are the exception to public records: they are never published row by row, only as counts for addons that at least three installs report. Guild data is published only through an officer's upload, and named records only with that person's sharing choice; a guild's boss kills (guild_kill, below) are the exception, published through the guild pages once two of its members' installs report them. Player rows are a second exception, described below: they carry character names, and they are never published as part of the public records.

The player pages are a separate thing, and they are off until the owner turns them on. Amended 17 November 2026 by the maintainer (docs/design/specs/2026-11-17-blizzard-api-first.md): the players module is on in a new addon install, which says so once in chat and in its feature chooser, with how to turn it off; an install from before keeps its choice. A full name (name and surname) also lets the armory find that character on Blizzard's API. When the players module is on in the game, the addon records one row per character: its name, realm, race, class, level and guild, and the day it was last seen. It never records where a player was, their gear, talents, professions, gold, achievements, earlier levels or guilds, which install reported the row, the character's id, or a time of day. The table holds one row per character, matched case-insensitively, and a newer sighting replaces the older one. A row is not corroborated: one upload makes a page, and the page says the data is unchecked. Whether a page should wait for more than one reporter is a question the lawyer review is answering (the player-pages design, section 13, question 7), because counting the reporters would mean keeping who reported a row. A row is deleted 90 days after the character was last seen, and a hidden character is deleted at once and never recorded again. The pages are unverified information reported by other players, and Everloot makes no statement that a character belongs to anyone.

The guild pages are off until the owner turns them on, like the player pages (the guild design, docs/design/specs/2026-10-24-guild-board-design.md). While the players module is on, the addon also records your own guild's boss kills: when ENCOUNTER_END reports a kill inside a dungeon or raid and the game shows your character in a guild, one guild_kill row with the guild's name, the realm, your faction, the encounter's id and name, the instance's map id and difficulty, the group size and the time. It holds no character name, no GUID and nothing about who else was there; at most 60 are written a UTC day, and /everloot players off deletes them with the player rows. guild_kill rows are never published row by row: they are not in the public corpus API or in corpus.ndjson, so one install's report and a hidden guild's rows never leave. A kill is the guild's once two installs report the same guild and boss, timed at the earliest of their reports; the guild pages, the guild tools and the open data export's guild-kills.ndjson and guilds.ndjson hold only those, and only while the guild pages are on. A guild's page lists its agreed kills, its officers' publication when they made one, and the characters the addon saw in that guild: these are player rows, each already public on its own page, grouped by the guild the game showed, and the page says they are sightings, not the game's member list. A character's page shows the guild the addon last saw it in as that guild's page shows it, and what the guild's officers published about that one character: its rank on their roster, the loot they gave it and whether it was at their last raid. A hidden character is left out of all of it. Anyone can hide a guild with the guild hide form, without an account or proof of office: the guild leaves the pages, the ladder, the search and the export at once, and its kill rows are dropped at intake from then on. Its characters keep their own pages, which the character hide form governs.

Boss pulls and runs. Inside dungeons and raids the addon records how long each boss pull took, kill or wipe (startedAt, durationSeconds and runId on the encounter row), and each run (instance_run): when you entered and left, the first pull, the last kill, how many pulls, kills and wipes, and which bosses died. No names. Both are on with the World switch. Runs are stored with the upload and not published yet: they wait for a rollup with its own agreement rule.

Group numbers in boss pulls. After each boss pull the addon also records one pull_summary row: the boss, kill or wipe and how long it took, and each member of your group by name and realm, with their class, their role and the damage done, healing done, damage taken and deaths the game's own damage meter (C_DamageMeter) showed. The role is the group role the game shows (assigned) or, without one, a guess from healing and damage (inferred: healer or damage, never tank). It never reads their gear, talents, guild, position or chat. Whoever uploads sends the whole group's numbers, and the protection is per character: anyone can hide a character with the character hide form, without an account. A hidden character is deleted from every stored pull at once and left out of every later upload, and a pull left with no member is deleted. The module is on in the addon, which says so once in chat; /everloot pulls off stops it and deletes the rows from the file. The rows leave the computer from the upload page, after a preview that counts them and says what they hold, or from the Desktop App; the command-line uploader never sends them. The site stores them in their own tables, never in the upload object or the public corpus, keeps each one for a year after it was received, and shows them nowhere: not on a page, in the corpus API, the open data export or the MCP tools. Any later display waits on a published policy for showing players and leaves hidden characters out. Nothing is collected until both the addon release that carries it and the site's switch for it are on.

The addon also reads a few game facts about the NPC you target and the quest you accept, as the game shows them to you. For an NPC: its creature type (Beast, Undead, Humanoid) and its rank (normal, elite, rare, rare elite, boss). For a quest: the quest log header it sits under (a zone or a category), the suggested group size and tag (group, dungeon, raid, PvP) when the game gives one, the NPC who gave it, and the faction of the character that accepted it, Alliance or Horde, never the character's name. These are facts about the NPC or the quest, not about the player, and the site never shows a faction next to a person. Each is published only when three installs report the same value. A quest's side is shown as Alliance or Horde when three installs on that faction report it and none on the other; it is shown as Both when three installs on each faction report it; otherwise it stays blank, and it is never guessed from a title, a giver or a zone. Nothing is read from the game's Creature or quest tables for these columns.

Places on the map come from where players stood, at two moments (addon 0.3.10). When an NPC's window opens (merchant, gossip, quest, trainer, flight map, bank, auction house, stable, innkeeper), the player is within a few yards of that NPC, so the addon records the player's position as the NPC's (positionKind: "interaction") with the service the window offered (vendor, trainer, quest giver and so on), once per NPC, service and zone. When the player opens a herb, ore vein, chest or other object, its loot row carries the player's position, how it was opened (from the spell cast on it) and the object's name as the game showed it, taken only from a herb, ore or chest cast, so the target of any other spell (a player you healed, say) never becomes a name. Where you stand when you only target or mouse over an NPC (positionKind: "player") is never used as the NPC's place. A Polestar Recorder file's NPC windows become the same interaction positions, and its kills and rares seen on the unit become player ones. A rare or treasure the minimap showed is at its own position, as the game's vignette API gives it (vignette for a rare, an object_sighting for a treasure), so it can make a pin. An NPC pin needs three installs on the same map; a treasure's, three installs within one cluster, as below; a dungeon entrance players zoned in from, three installs on the same map, and it confirms the entrance the game files place rather than adding a second one. A boss is listed for an instance once three installs report its encounter there. Instance, encounter and object names arrive in the player's client language, so each is shown only when three installs report the same one; an object's spawn point needs three installs within one cluster (positions closer than 1 on the 0 to 100 map scale), and a place seen by one or two installs is only counted, never drawn. An object's name is shown only when three installs report the same one; otherwise it is "Object" and its id. Fishing gives the zone only, never a pin. Object places come from play only: nothing is read from the game's files for them.

Item numbers players report. Proposed 2026-10-21, in the pull request that adds the item_fact kind (the maintainer review). The numbers on an item's tooltip that the game files do not hold (armour, block, damage per second, each stat and each resistance) come from players' own tooltips (addon 0.3.10). When the game shows the player an item's tooltip, the addon reads those numbers for the plain item, never an enchant or a random suffix on the player's copy, and records the item id, which line, the number and the client build. Nothing about the character is recorded, and nothing is read from a tooltip about anything other than an item. These are player data like any other upload: public as anonymous records, under CC BY-SA 4.0, and stated on a page as confirmed only when three installs report the same number for the same line. One or two agreeing installs show an unconfirmed value, labelled with the player count and a link to /addon. A patch can change an item, so the number is decided within the newest build that three installs reported. Each number shown says how many installs reported it, from which build and when it was last seen. A line installs disagree on is not shown; it stays [?], and the tooltip says the installs disagree. The game files' tooltip carries these numbers only where it is already shown (the item tooltip rule below).

Item pages may show how many distinct installs reported each tooltip line in its newest reported build when fewer than three have reported it (for example, "1 of 3 installs reported"). These progress counts are omitted for a line already agreed or contested in any build, include no install identity, and do not lower the confirmation threshold. The tooltip shows the partial value under the 2026-11-10 decision above.

Full auction scans are not public records either. They feed market values only, and a value is published when three distinct reporters scanned the item. For that count, an upload containing a scan is stored with a keyed hash of the address it came from (never the address itself). The hash stays in private storage and is never published. A scan also records the scanning character's faction (Alliance or Horde) and which auction house it read (Alliance, Horde or neutral), so each house can have its own market values, published under the same three-reporter rule; no character, seller or buyer is recorded.

When a player-started full scan cannot use replication, a bounded browse/search fallback may record actual listing prices and quantities as partial observations. Browse-summary minimums are never used to invent a complete market snapshot or listing distribution. The fallback records no seller or buyer information.

The addon's gold-per-hour sessions (/everloot gph) are personal: they stay in your SavedVariables file, outside the records an upload reads, and are never sent to the site.

EverlootDB.inventory is also local-only: explicit consent, off by default, stores item counts and observation times for bags, opened bank and opened mail under an opaque character key. It never enters upload records or private sync. Turning capture off deletes all inventory; forgetting a character deletes its stock. Partial or unreadable observations retain the prior dated snapshot, while a complete empty observation replaces it.

The addon's window keeps two more tables that are local in the same way. EverlootDB.ledger holds your money events, your auction sales and each character's closing gold per day, and EverlootDB.roster holds one row per character you log in with. Neither is a record, neither is read by an upload, and neither is sent to the site. A money event keeps at (its time), key (the opaque character key), category, copper and, when there is one, itemName and count, plus ownAlt when the gold went to or came from one of the player's own characters. An auction sale is read from the mailbox invoice and keeps itemName, count, unit, total (after the cut), deposit and cut when the letter carries them, and status (pending from the game's sold notice until the letter is collected, then paid; unmatched when no letter came within 48 hours), with its time and the character key; the buyer's name is never read into the file. To mark a trade or a letter as gold moved between the player's own characters, the addon reads the trade partner's name, the name a letter is sent to and the sender of a letter with money, compares each with the player's own characters on the same realm, and drops it: none of these names is ever kept. The completion tracker's rows are unchanged.

Completion tracking (addon 0.3.3, off until /everloot completion on) follows the same split. Four catalog kinds (achievement_catalog, reputation_catalog, title_catalog, quest_chain_catalog) hold only game facts the addon read from the game's own calls: ids, names, criteria text, standing headers and quest ids, with no player data. They reflect what one character's windows list, so they are not on the public allowlist and never leave the site raw: the pages (the collections tool, and the database's achievement, faction, title and quest chain lists and pages with their sitemap and search entries) and the MCP tool list an entry only after three reporters (distinct upload addresses, an IPv6 address counted by its /64 block, as the market counts them) report it identically; there is no copied catalog, and flavour text is not collected. An achievement's criteria wording is agreed separately: its page shows the list only when three reporters sent the identical complete list (a cut-off or blank list is never counted), and the list never appears on the lists or in search. For that count an upload containing a catalog row is stored with the same keyed address hash as a scan, kept private and never published. Four owned kinds (achievement_owned, reputation_owned, title_owned, quest_owned) hold one character's own progress under an opaque 8-digit key derived from the install id and the character's GUID, never the name. They are named data in the collection sharing category, like collection_owned, and are shown only while a verified install grant for that category is live. All eight kinds are kept out of the public corpus and folded into a separate completion rollup (latest row per character and entry). That rollup is private storage, but it records which address hash and install created each catalog entry, so it links an address hash to an install id. The link lasts until the rollup is rebuilt (?rebuild=1), which drops everything from uploads that were deleted; deleting a player's uploads therefore takes a rebuild to remove the link. Turning the module off deletes every one of these rows from the file.

With the same switch the addon also keeps five kinds for the Desktop App: profession_owned (each profession's name, rank and max rank), lockout_owned (each raid the character is saved to this week: instance, difficulty, bosses killed of total, reset time) and talent_owned (the active talent build: tree, spec if known, points spent, talent node ids and ranks), stats_owned (the character sheet's attributes, attack power, hit, crit, haste, expertise, spell power, healing, armor, defense, dodge, parry and block, as the game's character window shows them) and pvp_owned (PvP rank, rank points toward the next rank, top rank, season, lifetime honorable kills), under the same opaque key and holding only the latest state. These kinds stay on the player's computer except for private sync (see Private sync and owner-published sections): every uploader holds them back, the site's intake has no schema for them and refuses them, and the Desktop App's Chat does not pass them to a model provider. Only the Desktop App's sync sends them, and only for a character the player linked. Turning the module off deletes them with the rest.

Amended 17 November 2026 by the maintainer (docs/design/specs/2026-11-17-blizzard-api-first.md): from the addon release after that date, the addon no longer records achievement_owned, reputation_owned, quest_owned, profession_owned, talent_owned, stats_owned or pvp_owned, nor gear_snapshot and talent_snapshot, since Blizzard's character API supplies them; that release deletes any it finds in the saved file. title_owned, lockout_owned, boss_kill_owned and the catalog kinds stay. The site keeps accepting the dropped kinds from older addons until a later protocol release says otherwise.

The Completion tracker also keeps boss_kill_owned: encounter id, instance map id, difficulty, first and last kill time and kill count, under the same opaque character key. It records successful ENCOUNTER_END events only while Completion is on. Outside an instance it accepts only encounter ids of the shipped world-boss catalog; whether Forever emits that event outdoors is unconfirmed, and /everloot status counts the world-boss kills seen. These rows stay on this computer: the upload page holds them back and public intake has no schema for them. Turning Completion off deletes them. Draft amendment for maintainer review: once enabled, the Desktop App may sync these kills to the signed-in owner only with that character's separate boss-kill consent; they remain private, revoking consent deletes the synced history, and public uploads still refuse them. No public kill data is inferred from this local history.

The census counts who plays what. It is on in a new install of the addon, which says so once in chat; an install from before the census keeps it off until /everloot census on. It leaves the player's computer only from the upload page or from the Desktop App, if the player allowed the census there. It records race, class, level and faction for your own characters as you log in, plus the realm name, and for players near you: ones you target, mouse over or group with, and ones the game shows a nameplate for (NAME_PLATE_UNIT_ADDED, so the player's own nameplate settings decide which). For other players it reads those four things only, never their name, guild, gear or position. The addon keeps one row per character per day. Each row carries the character's GUID, the game's id for it, only on the way to us: before anything is stored, the site replaces it with a keyed hash (HMAC-SHA256 under a server secret) and keeps the server id, the number inside the GUID that says which server the character is on. Without the secret nobody can turn a GUID seen in game into our hash. If the secret is not set, census rows are dropped and nothing from them is stored. So what is kept per character is the hash, the server id, race, class, level, faction, and the realm name for your own characters, stored with the upload's install id like every other row. Census rows (census_self, census_seen) are never published row by row and are kept out of the public corpus, like addon lists. One install counts at most 500 players seen and 50 own characters a day, so no one install can invent a crowd. Only counts are shown, and a count appears only when it covers at least 10 characters reported by at least 3 installs. Below that it reads "fewer than 10".

Blizzard hides some WoW: Forever item details until an item drops, and runs no public test realm. Game files show what exists; play shows what is real. Everloot keeps the two apart and says which is which.

What the game files contain. Status: the table data is not published yet, except the talent catalog, the map layer, the raid buff table, the client layer, the race table, the Legacy perk catalog, the faction, title and quest line tables and the build changes described below. A maintainer can read tables of a game's current build from the files Blizzard's launcher downloads from its public CDN, the same files every player receives, and compare two builds field by field. The source ledger lists the tables we read. Those snapshots and comparisons stay on the maintainer's machine. Encrypted blocks are skipped, never decrypted: no decryption keys are loaded. Nothing else read from the game files is in the dataset, the API or the MCP server; the client layer, the race table and the catalog index below are the game-file layers they carry, and answers carry the game-file facts described under Game-file facts in answers. On the site, only the talent catalog, the map layer, the raid buff table, the client layer, the race table, the build changes, the icons, tooltip text and frame art covered further down, and the item names, quest titles and item set and spell names covered further down may appear.

Publishing the table data comes later, as a separate step. Once it ships, every entry from game files will:

  • be labelled found in game files, unconfirmed: a record there does not mean it is, or will be, in the game;
  • be hidden by default when it could spoil content that has not been released, behind a spoiler switch;
  • be reviewed by a maintainer before it is shown as a discovery; nothing will be published automatically;
  • come only from unencrypted data. Content Blizzard encrypts stays out;
  • stay Blizzard's data: shown as facts with the build it came from, kept apart from our CC BY-SA data and not under our licence;
  • credit WoWDBDefs. The column names come from the WoWDBDefs table definitions by the WoWDBDefs contributors, which are under CC BY-SA 4.0, so every file that shows them names WoWDBDefs and links its licence. On the site, the credits section of /methodology (/methodology#credits) does, and every page links to it from the footer.

When players later see the same thing in the game, the entry will show both: found in the files and observed in play.

Blizzard's icons, tooltip text and window frame. Decided 2026-10-18: our tools may show Blizzard's own talent and spell icons, the spell and talent tooltip text, and interface frame textures that make a tool look like the in-game window. Wowhead and most calculators do the same. Neither this decision nor the extensions below cover Blizzard's marketing art (key art, marketing logos, press kit, wallpapers), which the site does not use. The game's own logo, read from its interface files, is a different thing and is covered by the second extension. The rules:

  • Only from unencrypted files in the public client, read the same way as above. Nothing that needs a key.
  • Extracted by a maintainer, kept out of git. A maintainer extracts them from the local game client and uploads them to Everloot's asset storage with a build-stamped manifest. They are never committed to the git repository.
  • Stamped with the build they came from, in the file name or manifest and in the credits section of /methodology.
  • Still Blizzard's. We do not relicense them. They sit outside our CC BY-SA 4.0 data and our AGPL-3.0 code, and the site says so on every page, in the footer credit line below. The licence notes of every release file that includes them say so too.
  • Credited site-wide: every page's footer carries one line, "Game data and art © Blizzard Entertainment · Everloot data CC BY-SA 4.0 · Item and quest names from the Warcraft Wiki (CC BY-SA 4.0) · Credits", whose last part links to /methodology#credits. That section carries the full credit, with the build: "Icons, tooltip text and interface art are © Blizzard Entertainment, Inc. Everloot is a fan project and is not affiliated with or endorsed by Blizzard." Together they stand in for a credit line on each page.
  • Pages only. The data downloads, the API and the MCP server carry ids and facts, plus a link to the page where the icon and text are shown. They do not carry icon files, tooltip text or frame textures in bulk. If you want Blizzard's art in your own project, take it from the game client yourself.
  • Taken down on request. If Blizzard asks us to remove any of it, a maintainer removes it promptly, from the site first and from later releases after that. Blizzard, or anyone speaking for it, can write to us at [email protected] or through everloot.app/contact.
  • Same labels and spoiler rules as other game file data. Presentation does not change them: an entry found only in the files still says found in game files, unconfirmed, and anything that could spoil unreleased content stays hidden, icon and text included, until the reader turns spoilers on.

Extension: item, class and profession art, item tooltip facts and map pictures. Decided 2026-10-18 (Carlos: "Approved, ship it"). The rules above also cover:

  • Icons beyond talents and spells: item icons, and class, talent tree, raid buff and profession icons, read from the client's tables the same way.
  • Item tooltip facts: what the item tables (ItemSparse, Item) give for an item: name, quality, item level, required level, binding, slot, armour or weapon kind, weapon speed, stat kinds (not amounts), flavour text and sell price. Shown on a page only for an item that 3 installs have reported (a drop, a vendor offer, an auction scan), never in bulk and never for an id someone types into a URL. The one exception is an item the item name rule below names: its name, quality, icon and these facts. The extracted facts cover every named item in the build, released or not, so they are kept in private storage that only the site's server reads, never in the public asset bucket. Proposed 2026-10-22, in the pull request that adds them (the maintainer review): the same rules cover the lines the game prints from the ids those tables hold: the item set's name (ItemSet), the Unique-Equipped category (ItemLimitCategory), the classes and races that may use the item (ChrClasses, ChrRaces, announced races only), the skill, spell and reputation it requires (SkillLine, SpellName, Faction), and its Use, Equip and Chance on hit lines (ItemEffect, with the spell's description resolved like the talents'). Stat amounts, armour and weapon damage stay [?], and the tooltip says why in plain words.
  • Map pictures: the game's own continent and zone map art from the map art tables, laid under our zone outlines. A held-back zone's picture shows only with spoilers on. A continent's own picture always shows (decided 2026-10-19): it is the original world's parchment map, with no names and no new land painted in.
  • The /database pages (decided 2026-10-18 by the maintainer): the icons of the spells an item is created by or used in (from the client layer's relations) beside those relations, and pet family icons (CreatureFamily.IconFileID) in the manifest. Item icons and names on these pages follow the item tooltip rule above: only for items 3 installs reported, never in bulk and never for every id, except where the item name rule below allows them.
  • Dungeon and raid levels (proposed 2026-10-25, in the pull request that adds them (the maintainer review)): the level range the files give each instance, read like the entrance points beside it from the unencrypted tables MapDifficulty, ContentTuning and LFGDungeons, published in the map layer's file and shown on /database/instances and each instance's page as a fact from the game files. The files give one level for most instances, not a range, and none for some; an instance they give none for shows a dash. Nothing is filled in from memory or another site.

Second extension: loading screens, race and faction emblems, spell tooltips and the game's logo. Decided 2026-10-18. Carlos: "Approved next: (1) loading-screen art as page banners (zones, dungeons, raids; e.g. map zone panels, raid builder, patch notes header), (2) race and faction emblems (the race x class table on /wow-forever, census later), (3) spell/ability icons with hover tooltips where spells are named (patch notes /changes, raid builder buffs, talent pages already have them), (4) a launch banner at the top of the home page: the game's WoW: Forever logo from the game's own files plus our existing countdown plus loading-screen art, with 'Not affiliated with or endorsed by Blizzard Entertainment' visible on the banner itself." And: "Glint (mascot) is NOT part of this batch." The rules above also cover:

  • Loading screens: the picture each continent, dungeon and raid shows while it loads (Map, LoadingScreens), as page banners and beside dungeon and raid entrances on the map. A dungeon or raid added after the original world (a later map id) has its picture shown only with spoilers on.
  • Race and faction emblems: the client's race icons and the Alliance and Horde emblems, named pieces of its interface atlases. Only the races Blizzard has announced for WoW: Forever are extracted.
  • Spell tooltips where a spell is named: the raid builder's buffs, and the talents and raid buffs a patch note names, with the same icon and tooltip text as the talent calculator.
  • The game's own WoW: Forever logo: the logo piece in the client's interface atlas (logo-wow-forever), read like the other pieces. It appears only on the home page's launch banner, and only with "Not affiliated with or endorsed by Blizzard Entertainment" visible on the banner itself; the game's name beside it is plain text. It is never used to suggest that Blizzard made, approves or supports Everloot. No logo in the files means no banner: Blizzard's marketing logos and art are still never used, not even as a stand-in.

Item names from the game files and the Warcraft Wiki. Decided 2026-10-19 by Carlos ("Wiki match only") and widened 2026-10-21 by Carlos: "the goal is to name all the items that exist". Every item the current build lists (Item) is named when a source names it, before players confirm it, on the item page and the category lists, with its quality and icon and, where the game files hold them, its tooltip facts (item level, required level, binding, sell price, weapon speed, stat kinds and flavour text). The name comes from:

  • the game files of the current build (ItemSparse), wherever they name the item. Where the item's page in the Warcraft Wiki dump of 2023-11-13 gives the same name, apart from case, spacing and the kind of apostrophe, the name also carries the wiki's credit; where the two differ, the game files' name shows, as this build has it;
  • else the Warcraft Wiki dump's name, where only the wiki names the item.

The rules:

  • Each name says where it comes from in a tooltip on the name (and in words for screen readers; on a link, the tooltip is the link's hover card). A name the wiki gives reads Name from the Warcraft Wiki; the site-wide footer credits the Warcraft Wiki under CC BY-SA 4.0 beside Blizzard's credit, and /methodology#credits names the Warcraft Wiki contributors. A name from the game files alone carries the quiet unconfirmed marker, From the game files, not yet confirmed by players. A name 3 installs agree on comes first, with the tooltip Name confirmed by players. No source label is printed beside a name.
  • There is no id range: an item from a later version, an unreleased one, or one the files call "Deprecated" or "[PH]" is named as the files name it. The marker and the build stamp say what the name is: found in game files, unconfirmed. Items whose name marks them as development leftovers (test, placeholder and retired items, decided by kb/src/dev-item-names.ts, added 2026-10-26) are left out of the category lists and their counts unless the reader asks for them, and their pages stay out of search and the sitemap; their pages, names and id lookups stay as before.
  • The tooltip facts follow the name (added 2026-10-20, at Carlos's request that the database show the game file data it already holds, as Wowhead does). Until players confirm the item, its page keeps the quiet unconfirmed marker on the icon, whose tooltip says so, and prints no source line. The lists sort and filter by item level, required level and binding; an item whose facts are not shown matches none of those filters.
  • An item no source names keeps "Item N". The game server sends those names while the game runs, so they come from the addon's uploads, after 3 installs agree. A category page where fewer than half the rows have a name stays out of search.
  • Pages only. The data downloads, the API and the MCP server carry no item names and no tooltip facts for these items, as before; a test checks that their code does not read the names. The embeddable cards keep the three-install rule.
  • The names come from a ledger a maintainer builds with npm run kb -- item-names. It records every name each source gives for each item id, mismatches and single-source names included, and it only grows: a new build, dump or source adds to it and removes nothing. It stays on the maintainer's machine like the snapshots. The names the site shows go to the private storage the site's server reads, never to git or the public asset bucket. In build 1.60.1.70450: 31,849 items, 23,096 named (9,305 where the files and the wiki agree, 228 where they differ, 9,698 from the files alone, 3,865 from the wiki alone) and 8,753 left to players.
  • The maintainer review is the pull request that sets or changes the rule or adds a source. A file for a new build is checked against the last run's counts (npm run kb -- name-gaps) before it is uploaded.

Third extension: game icons and tooltip text in embeddable cards. Decided 2026-10-19 (Carlos approved it with the site structure plan). The hover card any site can show through https://everloot.app/tooltips.js and GET /api/tooltip/<type>/<id> may carry one item's game icon and tooltip facts, the same ones its page shows:

  • One item per request, and only for an item its page would show them for (3 installs reported it). Never in bulk, never for an id nobody reported.
  • Labelled and credited in the card: found in game files, unconfirmed, the build, and the Blizzard credit above. The icon is a link to our asset storage, not a copy in the response.
  • Same spoiler and takedown rules as the pages. The downloads, the rest of the API and the MCP server still carry no icons or tooltip text.

Quest titles from the game files and the Warcraft Wiki. Proposed 2026-10-20; decided 2026-10-21 by Carlos, and widened the same day to every quest the wiki titles. The game files list quests by id (QuestV2) but carry no titles. A quest's title may be shown on its page, in the quest lists and where an item page names the quest as a reward, before players confirm it, when the quest id is in the current build's QuestV2 table and the quest's page in the Warcraft Wiki dump of 2023-11-13 gives that id in its quest infobox. The title is the infobox's name, or the page title where the infobox has none, kept as the wiki writes it; a few carry the wiki's own note, such as "(Alliance, 55)". The rules:

  • As with item names, a title 3 installs agree on comes first. A wiki title says where it comes from in a tooltip on the title, Name from the Warcraft Wiki (and in words for screen readers; on a link, in the link's hover card); on the quest page it also carries the quiet unconfirmed marker. No source label is printed beside it. The footer credit reads "Item and quest names from the Warcraft Wiki (CC BY-SA 4.0)".
  • Any quest id the build lists, with no id range. Where wiki pages give one id different titles (a quest renamed in a later version, say), the first page's title shows and the run lists the id for review. A quest the wiki lacks keeps "Quest N" until players title it, and a wiki title for an id the build does not list is never shown.
  • Where a Forever-era source names a quest differently from the wiki, the wiki title is held (CONTESTED_QUEST_IDS, 9 quests found on 2026-10-21 against Polestar's guides) until players confirm one. The other source's title is never shown.
  • QuestV2 lists only quests with a completion bit (every row has a UniqueBitFlag), so repeatable quests such as Additional Runecloth or the Scourgestone turn-ins are not in it and never get a wiki title. Their pages and titles come from player uploads alone.
  • Pages only. The data downloads, the API and the MCP server carry no quest titles from this file; a test checks that their code does not read it.
  • A maintainer makes the file with npm run kb -- game-names, into the same private storage as the item names, never git or the public asset bucket. Unlike the item names it keeps no ledger: it is rebuilt from the snapshot and the dump each run. In build 1.60.1.70450: 6,614 quest ids, 3,229 titled (47 of them where the pages disagree), 9 held as contested and 3,376 the wiki lacks.

Item set and spell names from the game files. Proposed 2026-10-20; decided 2026-10-21 by Carlos, and widened the same day to every set and spell the files name. The names of item sets (ItemSet.Name_lang) and spells (SpellName.Name_lang) may be shown on their pages and beside the relations on the item pages. They come from the game files alone, so:

  • Each carries the quiet unconfirmed marker, with From the game files, not yet confirmed by players as its tooltip. No source label is printed.
  • Every set and spell of the build with a name, with no id range and no placeholder filter, as with items. One without a name keeps "item set N" or "spell N".
  • Pages only, in the same private storage, made by the same command. In build 1.60.1.70450: 536 sets, all named; 31,939 spells, 31,926 named.

The talent catalog. The talent calculator at /tools/talents uses our own catalog, read from the public client by a maintainer, not taken from any other site. The catalog is the trait node ids, positions, ranks, gates, prerequisites, spell ids, talent names and the build stamp. This is one of three table data exceptions to the status above, with the map layer and the raid buff table below. It may be published on the site, and it may be committed to the git repository as a clearly marked game-data file with a README that says it is Blizzard's and outside our licence. It is labelled found in game files, unconfirmed until someone has checked it in the game, and the spoiler rules above apply. Tooltip text, icons and frame art shown with it keep the rules above: asset storage, not git.

The Legacy perk catalog. Decided 2026-10-27 by the maintainer: the catalog of WoW: Forever's Legacy perks comes from the game files only, the Legacy trait trees read from the public client by a maintainer. No other site, guide or post supplies a value, and nothing third-party ships in it. It holds the tree, node and entry ids, positions, ranks, prerequisites, spend gates, the cost and the cap, with spell ids, perk names and the build stamp, and no descriptions, icons or tooltip text. Like the talent catalog it may be committed to the git repository as a clearly marked game-data file (web/src/data/game-data/legacy-perks.json, see LEGACY_PERKS), labelled found in game files, unconfirmed: it is Blizzard's data, kept apart from our CC BY-SA data and outside our licence. Each rule says how it was read, and a rule the files do not settle has no value. Decided 2026-11-02 by Carlos: a read-only page, /tools/legacy, lists it. No planner uses it: a Legacy planner waits until the cost, the cap and the spend gates it would enforce are confirmed in game or by Blizzard. It is not in the dataset or the API. Decided 2026-11-03 by Carlos: its perk names and ids reach the MCP knowledge search through the catalog index (see Game-file facts in answers).

The map layer. From the map tables we publish zone rectangles (UiMapAssignment), continents and instances (Map, UiMap), dungeon and raid entrance points (Map.Corpse) and, for each entrance and flight master, the zone the files put it in (for an entrance, the outdoor AreaTable area named after the instance when there is one, else the terrain chunk's area in the ADT files; either way up AreaTable to the area a zone's UiMapAssignment names; only the zone id is published), each stamped with the build it came from. Like the talent catalog, it may be shown on the site and committed as a clearly marked game-data file in web/src/data/game-data/, with a README that says it is Blizzard's and outside our licence. It is labelled found in game files, unconfirmed until checked in game. General rule: new zones or other content found only in the game files stay hidden unless the reader turns spoilers on. A zone that looks unreleased is held back until it is live: the file keeps it apart (heldBackZones), and /map draws it, dashed and labelled found in game files, unconfirmed, only when the reader turns the spoiler switch on (off by default). The public MCP tool leaves held-back zones out altogether. The map layer itself carries no art. The game's map pictures are part of the extension above: never committed, and kept in asset storage like the other Blizzard art.

The raid buff table. The raid builder at /tools/raid and its in-game view use our own table of group buffs and debuffs, read from the spell tables in the public client. For each effect it holds the effect id, its name, the class, the spell ids, who it reaches (party, raid, target), per rank where ranks differ, and the talent tree it needs where the game files prove that link, each with the build stamp. It holds no tooltip text. Like the talent catalog, it may be shown on the site and committed as a clearly marked game-data file in web/src/data/game-data/, with a README that says it is Blizzard's and outside our licence, and the addon may carry the same facts. It is labelled found in game files, unconfirmed until checked in game, and the spoiler rules apply.

The client layer. Decided 2026-10-18 by the maintainer: item categories (class, subclass, slot and quality) and relations read from the game files may be published on the site and through the MCP server, the API and the bulk download. The relations are created by (a spell's create-item effect, SpellEffect), reagent for (SpellReagents), part of set (ItemSet), same icon (items that share an icon file, Item.IconFileDataID) and, since 2026-10-18, profession (the skill line that teaches a spell, SkillLineAbility and SkillLine, with the profession's name). Profession planning also uses numeric teaching-item links (ItemEffect learning trigger 6 or base-difficulty SpellEffect learn-spell effect 36, joined through ItemXItemEffect), with no item or spell names in git. They are facts: ids and the relation between them, with no Blizzard text and no art. The rules:

  • Every entry is labelled found in game files, unconfirmed and stamped with the build it came from, on the page and in every answer. On a page the label is written once per section, not on every row; in an answer it travels with the rows.
  • It is Blizzard's data, kept apart from our CC BY-SA data: an API answer gives it its own notice (gameData) beside our attribution, and the file carries its own credit, WoWDBDefs included.
  • It is committed as one game-data file, web/src/data/game-data/client-layer.json, regenerated by the build watch with each build, so every change to it goes through a reviewed pull request.
  • Relations players see in the game (dropped by, sold by, reward from, located in) keep the three-install rule. Where both exist, a page shows both, each with its evidence.
  • Forever has no live build yet, so every entry comes from a pre-release build. Entries are not marked one by one as spoilers: the layer holds ids and relations only, no names, text or art. The item names a page may show beside it follow the item name rule above and are not part of the layer.
  • How it sits with the item tooltip facts above: class, subclass, slot and quality are public for every item through this layer. The rest of an item's tooltip (name, item level, required level, binding, weapon speed, stat kinds, flavour text, sell price) and its icon stay in private storage and appear only for items 3 installs reported, as the extension says, or under the item name rule above.

Races, classes and racial traits. Decided 2026-10-27 (Carlos: "proceed"), as an extension of the game-file layers above: which races can be which class, on which faction, and each race's racial traits, read from ChrRaces, CharBaseInfo, ChrClasses and ChrRaceRacialAbility. The rules:

  • Race and class ids, the race and class pairs (CharBaseInfo), each race's faction and the trait names (with active or passive and their order) may be published on the site and through the API, the MCP server (the class_races tool) and the bulk download.
  • A trait's description is game text. It stays on pages, the same rule as tooltip text: the class pages and /wow-forever may show it, and no API answer, MCP answer or bulk file carries it.
  • Every entry is labelled found in game files, unconfirmed and stamped with the build, once per section on a page and with every answer, and the file credits WoWDBDefs, like the client layer.
  • It is committed as one game-data file, web/src/data/game-data/races.json, written by npm run kb -- client-races and regenerated by the build watch, never edited by hand. A test compares its pairs with Blizzard's race and class article and names any pair the two disagree on; neither is preferred.
  • Skyborne, the new race, is announced by Blizzard in its own posts, so its traits are not marked as spoilers.

Game-file facts in answers. Proposed and decided 2026-10-29 by Carlos ("approve"). answer_sources (the site's chat, the MCP server and in-game /ask) answers from the committed game-data files when a question names an instance, boss, zone, flight master, mount, achievement or talent the 2023 wiki and player data do not cover. The rules:

  • An answer may carry the names those files hold for instances, encounters (bosses), zones, flight masters, mounts, achievements and talents, with ids and facts: an instance's kind, levels, zone and encounters in order, a mount's riding skill and spell, an achievement's points and criteria count, a talent's class, tree and ranks.
  • Game text stays on pages: achievement descriptions and criteria, spell and talent descriptions and tooltip text are in no answer, as for the race traits above.
  • Only for a question that names the thing: one fact per name asked, at most the answer's limit. The bulk download and the rest of the API carry none of it.
  • Every fact is labelled found in game files, unconfirmed and stamped with the build, with the game-data credit (WoWDBDefs included) beside it.
  • Held-back zones stay out altogether, and so does anything in them (a flight master there is not named), as the map rule above says.
  • The facts come from the files the build watch regenerates (web/src/data/game-data), so every change to them goes through a reviewed pull request.

Amendment, decided 2026-11-03 by Carlos: the catalog kinds (achievements, factions, titles, quest lines, mounts, legacy perks, races, instances) reach the MCP knowledge search through the catalog index, as names, ids and the listed fields, each labelled found in game files, unconfirmed with its build; achievement descriptions and criteria text stay on pages. The index is built from the same committed files, outside the CC BY-SA knowledge release, and search_knowledge, search, get_by_name, get_entity and fetch read it beside the release. Every such answer carries the label and the build, the game-data credit, and a link to a page that exists on the site: only the factions the game shows are indexed, and statistics are not achievements.

Build changes, beta builds included. Decided 2026-10-18, the same way Wowhead and other datamining sites work: when Blizzard ships a new build, a reviewed pull request records what changed in the tables we read, and that change list may be public in the PR text and committed in web/src/data/game-data/changes/. This includes builds that are not live yet, such as the beta, and the numbers in the files: spell effect values, durations and the like. Every entry carries its build and is labelled found in game files, unconfirmed. Anything that could spoil unreleased content is marked as a spoiler and hidden on the site unless the reader turns spoilers on; the repository itself is public, so the marking is a warning, not a lock. Encrypted content stays out, because we never decrypt. Description and tooltip text is not part of these files.

Blizzard's own posts. The "what changed" page at /changes lists posts by Blizzard staff on the WoW forum and by the official World of Warcraft account on Bluesky, and the game builds Blizzard's version service reports.

  • Excerpts only. For each post we keep the title, the author and role, the time, the category, an excerpt of at most two sentences (300 characters) and a link to the original. The text is Blizzard's, so we never copy a post in full. Each post has its own page on Everloot (/changes/<id>-<slug>) with those fields, the excerpt marked as Blizzard's words, and the link. A maintainer may add a write-up in our own words on that page (kb/curated/blue-writeups/), reviewed before it is published; it never repeats more of the post than the excerpt, and nothing writes one automatically. A page without a write-up is not indexed by search engines.
  • Sources. Forum posts come from the US and EU English WoW forums, only from the WoW: Forever and WoW Classic categories, and only from Blizzard staff: people on each forum's published staff list (about.json) or posts the forum itself marks as Blizzard's (its group tracker; the green MVP groups are players and are left out). kb/curated/blue-accounts.json holds hand corrections. Players' posts are read with a topic but never kept. Bluesky posts come only from the verified official World of Warcraft account, and only when they are about Forever or Classic. No third-party tracker is read or copied.
  • Robots and manners. The reader checks robots.txt before every run and uses public JSON only. Blizzard's forum .rss routes are disallowed by its robots.txt, so we never use them. It names itself (Everloot/1.0, with a contact address), never follows a redirect, and stops when a server answers 429 or a server error, honouring Retry-After.
  • Polling cap. At most once every 30 minutes and at most 40 requests a run in all: the staff list at most once a day, the Forever and Classic topic lists only back to what the last run saw, and only topics that changed and carry a staff post. A source asked less than 25 minutes ago is skipped. A one-off history import (blue-backfill) uses the same rules and pace with a larger budget.
  • Build times. A build's time on that page is when Everloot first saw it, not when Blizzard released it.

Fan permissions are narrower than this. As we read Blizzard's video policy, its permissions for fan use of game material are non-commercial and Blizzard can revoke them. The older Blizzard fan-site FAQ is only quoted secondhand to us, so we treat what it says as unverified and rely on nothing from it. Since PR #62 (2026-10-17) Everloot shows clearly labelled Google ads on the tools, the patch notes, the home page and database pages that have content, and none on the upload page, the privacy page, account pages or the lore archive; the privacy policy and the terms describe them. Some of those pages show Blizzard's icons and tooltip text. Whether that sits within Blizzard's current terms is the operator's call to review; if the review says no, the ad slots come off those pages, not the assets.

The addon and project tracker. /addons/ecosystem lists WoW: Forever addons and tools that a maintainer has checked and added to kb/curated/ecosystem.json.

  • GitHub facts only. For each project we read public repository facts from GitHub's REST API: stars, latest release, last commit, commit and issue counts, a contributor count, and the ## Interface line of the project's .toc. No usernames, emails or contributor lists are kept or shown, and no person is ranked. The boards rank projects.
  • Polite and slow. Serial requests, at most every three hours per project, an optional token, and a stop (never a retry storm) when GitHub says to wait.
  • Installs. A project's install count comes from our own addon-list uploads and appears from three installs up, like /addons/popular. A project with fewer shows "not counted", never zero.
  • Other hosts are link only. CurseForge's API terms forbid caching its data, Wago documents no read API for listings and WoWInterface's terms forbid automated access, so we read none of them.
  • Opt-out. An author can ask to be removed with the ecosystem issue form or at [email protected].

Addon pages and the usage board. Two surfaces use addon names, and only one of them publishes a catalog entry.

  • The usage board counts. /addons/popular lists an addon once three installs report it, with its share, as an anonymous aggregate. For an addon nobody approved, the title is the plain title installs agree on and labels the count only: the row has no page, no download links and no repository facts.
  • The catalog is approved by the owner. An addon gets a page at /addons/<name>, a title, repository facts and download links only when it is named in the curated source map (web/src/data/addon-repos.ts) or by a project in kb/curated/ecosystem.json. Both change through a reviewed commit, never through an upload. An unapproved addon answers 404 however many installs report it, the page reads nothing from the upload store to decide that, and no upload field (title, name spelling, version) reaches an approved page's title, link or source. Install counts are joined to an approved page as usage only.
  • Approving one. Add it to the registry (or the curated map) with its repository and the page that shows it is for Forever. The deploy that ships the change starts a new edge cache version for the addon pages, so a page built from the old list is not served after it.

Open sources. The lore archive is built from the Warcraft Wiki dump of 2023-11-13 (CC BY-SA 4.0). We may add other openly licensed datasets later. Each one goes in the source ledger with its licence and exact version before any of it ships. A Classic 1.12 baseline, pfQuest's database (built from the VMaNGOS and CMaNGOS emulator databases, GPL-2.0 and GPL-3.0), is used on a maintainer's machine to check what Forever players have confirmed and where they have not looked yet. It is never published and never mixed into the CC BY-SA data: only what players' uploads confirm is published. Decided 2026-10-21 by Carlos: the coverage page, the API and the MCP tool may show counts and percentages whose denominator comes from the baseline (how many facts of a kind a zone is expected to hold), but never a baseline name, id or row.

Facts we write ourselves, in our own words, each with the page it came from and the date.

Guides. A guide (kb/guides/, contract) cites only Blizzard's own posts, on worldofwarcraft.blizzard.com, news.blizzard.com and the US and EU forums, each linked where its claim is made with the day it was checked, and Everloot's own data through its pages on the site. It adds no source: other sites' guides are not sources, and the ledger's exceptions stay in the files and pages they are scoped to. npm run kb -- guide-check refuses any other link off the site.

Game file data, and the icons, text and art shown on our pages, stay Blizzard's and are not under our licence. Everything else Everloot publishes is under CC BY-SA 4.0: anyone may reuse it with credit, and what they build from it must stay under the same licence.

Recently public facts

The database's recent-facts page and RSS feed read a snapshot written every ten minutes. They list at most 500 entities from the last 30 days: agreed item names or public rates, NPC pins or loot, quest givers, object loot, bosses and entrances, under the same publication rules as their pages. Entries carry only the entity's kind, id, public name, page link, agreement count and publication time. No install ids or player names are included. The snapshot also keeps the current and ever-public entity keys (kind and id only, capped at 100,000) in a separate inventory read only by the cron, so an old capture uploaded today is dated when it first becomes public. Existing entities on the initial backfill have estimated dates from report capture days; the page and feed label those estimates. A publication-rule version change establishes a new baseline without announcing existing facts. A failed snapshot write is retried.

What we refuse

  • Bulk scraping of any site. No crawlers, no copying another database's rows, guides or text. If a site's terms or robots.txt say no, the answer is no.
  • Touching the running game. No reading or changing the game's memory, no automation, no bypassing protections. The addon uses only the game's own addon API. The Desktop App and the command-line uploader read the SavedVariables file the game writes, and writes /ask answers into a file the addon loads after /reload. The upload page reads only the file a player chooses: Everloot's own, or a Polestar Recorder file the player recorded and chose to upload.
  • Decrypting game data. Content Blizzard encrypts stays encrypted.
  • Bulk exports of Blizzard's art or text. Icons, tooltip text and frame textures appear on pages only, never in the downloads, API or MCP server.
  • Unreviewed spoilers. No game file entry is shown as a discovery without review, and anything that could spoil unreleased content stays hidden, icons and text included, unless the reader turns spoilers on.

Blizzard's EULA forbids unauthorized software that "intercepts, collects, reads, or 'mines' information generated or stored by the Platform." Studying released game files offline is a long community practice that Blizzard has tolerated, and we read only what its CDN serves to every player. If Blizzard asks us to stop, we will.

Wowhead

We link to Wowhead and never copy it. Looking up single facts there is allowed only if every one of these holds:

  1. Nothing is built until the terms allow it. A maintainer reads Wowhead's Terms of Use and confirms that automated lookups of single pages are not prohibited, or Wowhead gives permission. If the terms forbid it, the feature is dropped. As of this writing, it has not been built.
  2. Facts only: ids, levels, coordinates and looted counts. No text, images or guides.
  3. Gaps only: a lookup runs for an entity our own data lacks or disputes. No crawling, and never Wowhead's list or random pages.
  4. Honest and slow: the user agent is EverlootFacts/1.0 (+https://everloot.app/contact), never a browser string or a bot name Wowhead blocks. At most one request every few seconds, and every response is cached so a page is fetched once.
  5. Local only. The facts are kept apart, marked not publishable, and used only to check and fill gaps in our own observations. They are never republished on the site, in the dataset or by the MCP server unless Wowhead gives written permission.
  6. Run by a maintainer, never by an automated agent. The script lives in this repository, so anyone can read exactly what it does.

Contributing data

  • Name the source in your pull request and add it to the source ledger.
  • Only sources whose licence allows republishing under CC BY-SA 4.0 can reach a release. Anything else stays local and is marked not publishable; the release build refuses it.
  • The one exception is the game file layer described above. That data is Blizzard's, so it never enters the CC BY-SA release; once publishing ships, it reaches the site only as its own labelled layer, after a maintainer has reviewed it. Icons, tooltip text and frame textures are never committed to the git repository, so a pull request must not include them. A maintainer extracts them from the local game client and uploads them to Everloot's asset storage with a build-stamped manifest.
  • If you are unsure whether a source is allowed, open an issue first.

Using our data: credit, limits and keys

Everloot's data stays open. Anyone may read it, copy it and build on it under CC BY-SA 4.0. Two things are asked in return: credit it, and use a free key when you copy all of it.

  • Credit. Show this line, with the link, wherever the data appears: "Data from Everloot (https://everloot.app), CC BY-SA 4.0." What you build from it stays under the same licence. Every answer carries it so it is hard to lose by accident: the JSON APIs return it in an attribution field (licence, credit line, the page the data comes from) and in the X-Everloot-Licence, X-Everloot-Credit and Link headers, which every MCP response carries too. An MCP answer links the everloot.app page it comes from instead of ending with the licence line: the link is its credit, and an assistant that shows it has credited the data. Republishing the data without credit breaks the licence.
  • Asking questions is free and needs no key. The MCP server and the market prices keep a per-address limit that a person or an AI assistant answering questions does not reach.
  • Copying in bulk takes a free key. Once the limits are switched on, the corpus API and the drop list answer a few requests an hour per address without a key, one page of records at a time. An IPv6 address counts by its /64 block. A free key, made on /docs/api/keys after signing in with Battle.net, raises the limits far enough to copy everything. The allowance belongs to the account, across all its keys, and an account can make only a few keys a day, so new keys do not bring new allowances. Requests are counted per key, so we know who copies the data in bulk and can ask what they are building. Keys can be revoked or deleted by their owner at any time, and revoked by us when a key is used to republish the data without credit. The exact numbers are on /docs/api/keys.
  • Pages are not limited. Search engines and AI assistants may read and quote the site's pages. These limits apply only to the machine-readable endpoints.
  • Not covered. Game-file data and Blizzard's posts are Blizzard's, not ours; the credit line and the licence above do not apply to them. The code is AGPL-3.0 and has its own terms.

What we keep about keys, and for how long. For each key: a one-way hash of it, its last four characters, its name, what its owner says it is for, its permissions, its account, when it was made, when it expires, when it was last used and revoked, and a count of its requests per day and endpoint. Not what was asked. The purpose is asked when a key is made so we know what keys are used for; it is not published. A key made on /account expires on the date its owner picks (7, 30, 90 or 365 days, or a date up to a year ahead); past it, every endpoint refuses the key, as it does a revoked one. Keys made before expiry existed, and the Desktop App's key, have no end date. Daily counts are kept 90 days. A revoked key and its counts are deleted 30 days after it was revoked, or at once when its owner deletes it on /account; an expired key stays until its owner deletes it. Deleting an account (write to [email protected]) deletes its keys and counts. Anonymous requests are counted per hour against a keyed hash of the address, which is deleted after about a day; the address itself is never stored.

Saved builds

A signed-in player can save builds from the talent calculator to their account. This is the player's own data, not game data and not an observation, and none of it is published as a dataset.

  • What a saved build holds. The calculator's share codes for one or two talent sets (the same EL1-... code a link carries), the game build they were made on, a name, optional notes, an optional character label, and whether its short link is on. Nothing from the game files and nothing the addon observed.
  • Character labels are typed, not checked. A label is a name, a class, an optional realm and an optional note that the player types. It is not looked up in the game, the census or the player pages, and it is never shown to anyone else unless the player shows their characters on a public profile (below).
  • A shared link is public to anyone who has it. Each build has a short link, /b/<code>, on by default. While it is on, anyone with the link sees the build's name, notes and talents, but not the character label. The owner's name shows only if the owner chooses to show it. The page is not indexed by search engines. Turning sharing off or deleting the build stops the link at once.
  • Keys. A free key made with "Read my saved builds" (read:own) can read its owner's builds and labels, for the Desktop App, MCP clients and scripts. A key without it cannot.
  • Limits. 200 builds and 30 character labels per account, and 300 saves a day.
  • Deletion. The owner deletes a build or a label on /me at any time. Deleting a label keeps its builds, with no character. Deleting the account (write to [email protected]) deletes every saved build and label with it.

Profiles

A signed-in player can set up a profile on /settings. Like saved builds, this is the player's own data, not game data and not an observation.

  • What a profile holds. A handle (3 to 24 letters, numbers and hyphens, unique), an optional display name, an optional status (one emoji and up to 80 characters, with an optional time after which it stops showing), a main character and the character the player is on now (both chosen from their own character labels), and the public profile switch with what it shows. Nothing is looked up in the game, and no picture is stored: the avatar is drawn from the name's initials.
  • Private by default. Until the player turns the public profile on, only they see any of it. On, /u/<handle> shows the handle, the display name and only what the player ticks: characters (name, class, realm), shared builds (their short links) and the status. It never shows the BattleTag, the email, a label's note or a build that is not shared. It is indexed by search engines only while it is public; off, unknown or without a handle, the address is a 404.
  • Limits. A handle changes at most 5 times a day. Reserved handles (admin, everloot, support and a few more) cannot be taken.
  • Deletion. The player changes or clears any of it, or turns the public profile off, on /settings at any time. Deleting the account (write to [email protected]) deletes the profile.

Public character pages

A signed-in player who has linked one of their characters (Linked characters, on /account) can give it a public page at /characters/<ruleset>/<realm>/<name>. The player's part of that page holds the player's own data from their own addon file and nothing about anyone else. The same address also shows, in their own parts, what the player pages and guild pages hold about the character (see "What players see in the game").

  • What a page shows. The character's name, level, race, class, faction and realm, as the addon file sent them when the character was last linked, and the date of that upload, worded "From the player's own addon upload, ". Nothing in the player's part comes from the player pages, the census, other players' uploads or Blizzard. From 17 November 2026, when the owner associated the character with its armory entry, the page also shows that entry's gear, stats, talents, professions, PvP, achievements, reputations and completed quests in a part of its own, labelled "From Blizzard's API, " (see Armory). It follows the page's switches below.
  • What it never shows. The BattleTag, the email address, the account, a profile handle, API keys, the character's game id, or the hashes a link is stored under.
  • Off until the player turns it on. A new link is private. The player's part shows only while the player has "Public page" on for that character. Turning it off, unlinking the character or deleting the account removes that part and its share cards at once; the address then shows only what the player pages hold about the character, or answers 404 like any unknown character when they hold nothing.
  • A link is not proof. A link shows that the account holder also had an addon file in which the character typed a code. It is not checked with Blizzard, and the row can be written by hand. The page says the data is self-reported and not checked against Blizzard, and it never calls the character verified or owned.
  • Disputes. When a second account links the same character, the character is frozen on every account that holds it and has no public page until we settle it. To settle it, write to [email protected]: we reply within 7 days, and the character stays frozen until it is settled.
  • Hiding. Anyone can hide a character through the player-page hide form, without an account. A hidden character cannot be linked and has no public page, whoever linked it. Each character page links to the form.
  • Share cards. Each page has two images, 1200x630 for link previews and 1080x1080 to download, drawn from the same fields as the page. Text and shapes only: no game art and no Blizzard marks.
  • Search engines. Every character page asks search engines not to index it. Players cannot change this yet.
  • Caching. Everloot draws the page and cards fresh on each request and keeps no copy. A copy that a link-preview service (Discord, X and others) has already fetched is outside our control and may stay with that service.
  • One page, separate parts. A character's page shows the linked player's part first, then what other players' addons saw of the character, then the guild it was seen in, each under its own source line and date. Neither adds data to the other, and the share cards draw only on the player's part.

Private sync and owner-published sections

In effect from 6 November 2026, accepted by the maintainer that day (EVERLOOT_PRIVATE_SYNC, migration 20261106120000_private_sync.sql). The active talent build syncs as the addon's keyed talent_owned row, which names its character by the same opaque key as the other synced rows, under the talents switch; there is no separate category. The public talent_snapshot row carries no character and is never synced.

Amended 6 November 2026, accepted by the maintainer that day: the active talent build (talent_owned). Amended 7 November 2026: character-sheet stats and PvP standing (migration 20261107140000_private_sync_stats_pvp.sql); to be accepted by the maintainer in PR review.

Amended 17 November 2026 by the maintainer (docs/design/specs/2026-11-17-blizzard-api-first.md): Blizzard's character API now supplies gear, talents, professions, stats, PvP, achievements and reputations. The gear, talents, professions, stats, pvp and reputations categories are retired: the site stops accepting them and deletes what was synced for them. The achievements category keeps only titles, which the API does not return. Raids and boss kills stay. The list below describes the categories as they were before this amendment; the retired ones stop when the migration that deletes them is applied.

A signed-in player can sync their own linked characters' progress to their account and choose which parts of it their character's public page shows. Like saved builds and profiles, this is the player's own data, not an observation of anyone else, and none of it is published as a dataset.

  • What is synced. For each linked character the Desktop App sends the categories below, and the site keeps only the ones the player has turned on for that character: equipped gear (slot, item id, item level), the active talent build (talent node ids and ranks), professions (name and rank), reputations (faction id and standing), this week's raid lockouts (instance, difficulty, bosses killed, reset), achievements this character earned and titles held, the character sheet's stats (attributes, attack power, hit, crit, haste, expertise, spell power, healing, armor, defense, dodge, parry, block, as the game's character window shows them) and PvP standing (rank, rank points toward the next rank, top rank, season, lifetime honorable kills). Each category has its own switch and every category starts off. The Desktop App sends these with a key made with "Sync my linked characters" (sync:own); that key can send and nothing else. It cannot turn a category on, change what the page shows, or list anything.
  • The link decides, not the computer. A synced row is kept only when the account that sent it holds a live link to that character and has turned that category on. It may come from any computer. A row for a character the account has not linked, or whose link is frozen, is dropped. Another account's link never receives it.
  • Only the latest state is kept. For each character and category the site keeps one capture: the newest. The same capture sent again changes nothing, and an older one never replaces a newer one. No history is kept.
  • What is stored. The character's link, the category, the capture's time, the values above, and a keyed one-way hash of the install id that sent it, so everything one install sent can be found and deleted. The install id, the addon's character key and the character's name are not stored with a synced row.
  • Private by default. Synced data is shown only to the player, on a signed-in preview of the character. It appears on the public page only when the page is on and the player has also switched that section on: gear, stats, talents, PvP, professions, reputations, raids, and achievements and titles (under Collections), each on its own switch, all off by default. A frozen or hidden character shows none of it, to anyone. The page says the data is self-reported through the Everloot addon and not checked against Blizzard; it never says verified.
  • Names on the page. An item is named only when three installs agree on its name; otherwise it reads "Item ". Factions, achievements and titles are named only when the catalog's three-install rule agrees on them, and a faction or title it does not list is left out.
  • Deletion. Turning a category off deletes what was synced for it at once. Turning a section's switch off takes it off the public page on the next request. Unlinking the character or deleting the account (write to [email protected]) deletes everything synced for it at once. A copy a link-preview service has already fetched stays with that service.

Gear rankings

In effect from 6 November 2026, accepted by the maintainer that day (EVERLOOT_LEADERBOARDS, migration 20261107120000_gear_board.sql). With the switch off, /rankings/item-level and /rankings/gear-score answer 404 and the switch below is refused. The stored boards and the daily history below (migration 20261107150000_profile_pipeline.sql) are proposed: they take effect once the maintainer accepts them and applies that migration.

Amended 17 November 2026 by the maintainer: these synced boards give way to the armory boards (/armory/rankings), which rank every character the armory holds from Blizzard's data. /rankings/item-level and /rankings/gear-score redirect there once the site change ships, the synced gear category is retired (see Private sync), and the stored boards and their history are deleted with it. Until then the rules below hold.

Two public boards rank linked characters by the gear their owner synced: average item level and Everloot Gear Score. This is the ranking contract the owned-character identity ADR asks for, for gear only. Achievements, gold and the Everloot Score are not ranked by it.

  • Who is on them. Only a linked character whose owner made its page public, turned gear sync on and turned "Rank my gear" on (ranked_gear, off by default). Nobody else: no sighting, inspect, census row, other site or Blizzard data puts a character on a board. A frozen or hidden character is never on them. Gear older than 30 days is left off.
  • What is shown. The character's name, realm, level, race and class, its item level and Gear Score, its rank, and the date of the gear. The items themselves appear only on the character page, and only when the owner shows gear there. The character page shows its own rank on each board.
  • How the numbers are made. Both are averages over the same 17 equipment slots (shirt and tabard left out), an empty slot counting 0, a two-handed weapon filling both hands. Item level is the plain average. Gear Score v1 weighs each slot by its stat budget at item level 60 from the game's RandPropPoints table (the Epic column), read from the Forever client files and labelled found in game files, unconfirmed: 44 for head, chest and legs; 33 for shoulders, hands, waist, feet and trinkets; 25 for neck, back, wrists, rings and off hand; 18 for the main hand; 14 for the ranged slot. Which slot falls in which budget group is the long-documented classic mapping, not something the files state. These five numbers are the only thing taken from that table, and they are published in full on /methodology. A change to a weight is a new version.
  • Checked against the game files. Every item level counted is the game files' level for that item. Gear holding an item the files do not list, or one a page may not show yet (not reported by players and not named under the item name rule, so possibly unreleased), an item level that differs from the files, or an item in a slot it cannot go in, is not counted at all, and the board says how many characters were left out. The gear is still the player's own upload: the board says it is not checked with Blizzard.
  • What is stored, and for how long. Every 30 minutes the site scores the boards and keeps the result in one private file (the rows above, with each character's link id). Each page view filters that file against who is on the boards right now, so turning "Rank my gear", the page or gear sync off, a hide request or unlinking takes the character off on the next request, not at the next scoring. While a character is ranked, its level, item level and Gear Score are also kept as a history, at most one value of each per day (the day the gear was captured), and its character page draws that history as a chart once there are two days. The history is shown only while the character is ranked, kept while it is not, and deleted with the link: unlinking or deleting the account removes it, as it removes the gear.
  • Search engines. The boards are noindex until indexing of character pages is accepted.

Armory (Blizzard data)

The armory gives Forever characters and guilds their own pages from the Blizzard Developer API. These pages are separate from the addon sightings and the linked-character gear rankings above.

  • No ads. Armory pages carry no ad slots. The site's consent message may still appear.
  • Free, never in our dataset. Amended 17 November 2026 by the maintainer (docs/design/specs/2026-11-17-blizzard-api-first.md): armory data may also be shown on character pages, the Completionist tools, rankings, guild boards, MCP tools and API answers. Wherever it appears it is read from the armory tables, never copied into the addon tables, labelled "From Blizzard's API" with the time of the read and credited "Data from Blizzard Entertainment", and it disappears when the armory deletes it (the thirty days below, a 404 or a hide). An API or MCP answer carries it in its own field with its own notice. It never enters the CC BY-SA dataset, the public corpus, open-data snapshots or bulk exports. Pages that show it carry no ads.
  • Thirty days. Characters and guilds are checked again within 20 days. Stored data not refreshed for more than 30 days is deleted. A Blizzard response confirming that data has not changed (304) counts as a refresh. When Blizzard answers that a character or guild no longer exists (404), its stored row is deleted at once.
  • Credit. Every armory page says "Data from Blizzard Entertainment" and shows the time of the newest read. Nothing on it claims Blizzard endorses Everloot.
  • One client. The crawler uses one Blizzard API client. Its client id, client secret and the database service key stay in GitHub Actions secrets, never in a file, a test or a log.
  • Hiding. Anyone can hide a character at /armory/hide with its region, name and surname. Its character row, roster memberships and queued character requests are deleted. The hide is kept so later API reads and rosters cannot put the character back.

Blizzard auction prices

An hourly job reads every Forever auction house from the Blizzard Developer API and keeps one summary per item and house: the lowest unit buyout, the quantity-weighted median unit price, units and listings. Raw auctions are not stored. The job stays off until the maintainer sets the repository variable BLIZZARD_AUCTIONS to on.

  • Apart from player prices. The summaries have their own tables. They never enter the market rollup built from addon scans, the addon price file, the CC BY-SA dataset, open data, MCP tools, chat answers, API answers, sitemaps or exports. web/src/lib/blizzard-auctions.test.ts fails if another file reads them.
  • Labelled and credited, on a page of their own. /blizzard-auctions/<item id> shows them under "Blizzard auction house data, as of" the time Blizzard published them (each row also shows its own), with "Data from Blizzard Entertainment". The item page shows the player-reported prices and only links there; it reads no Blizzard data. A read that fails says so, apart from an item with no data yet.
  • No ads. AdSense loads on every other public page for the consent message, so the Blizzard page is an ad-free page (web/src/data/ad-free-paths.json): it loads no ads, analytics or consent banner, and a move to it from any page with ads loads a fresh document. It is noindex, disallowed in robots.txt and in no sitemap.
  • Thirty days. A house's items and commodities are two documents, each stored with its own publish time. Each item keeps one row per house, document and UTC day, its newest listing that day. Rows and houses older than 30 days are deleted every pass, and by a purge job that runs on schedule whether the refresh is on or not, with the Supabase secrets alone. The page never reads a row older than 30 days, and no shared cache keeps it. Backups skip these tables, since a backup outlives 30 days.

Chat

/chat answers questions about the game from Everloot's data. Everloot search needs no model and sends the question nowhere but this site. A player may add a model with their own key; that key stays in their browser tab and the page talks to their provider directly.

  • Everloot (free). When the site offers it, a visitor with no key of their own gets a built-in model on providers' free tiers: Groq first, then Cloudflare Workers AI when Groq is busy or at its limit, before any of the answer has been sent. The question, the earlier turns of that chat and what Everloot's tools returned go through this site's server to the provider, which writes the answer; each answer names the provider that wrote it. A character or note goes only if the visitor attaches it. No account data, no key, no cookie and no address goes to the provider. Each provider's key stays on the server, and a provider is never sent another's key.
  • What we keep. Not the questions or the answers. To enforce the limits the server keeps a keyed hash of the address, never the address, for a day. The providers' own terms cover their side: Groq's say it does not train models on what it is sent and keeps it at most 30 days, for reliability and abuse checks; Cloudflare's say it does not use Workers AI content to train models and does not keep it once the request is answered.
  • Limits. A visitor gets 6 model calls a minute and 45 a day, whichever provider answers; one answer is usually 2 or 3 calls. The site makes at most 900 calls a day to Groq and 55 to Cloudflare, each under its free tier. Past a limit, or when the free tiers are used up, the answer is Everloot search, with a line saying so.
  • Free means free. The model runs on the provider's free tier only; Everloot pays for no model.

The question ledger

What people ask Everloot is the signal that decides which data comes next (docs/design/specs/2026-10-27-question-bank-design.md). The ledger records that signal without recording the people or their questions.

  • What it holds. For each lookup made through /api/tools or an MCP mount (the site's chat, the Desktop App, in-game /ask, AI clients) with search_knowledge, get_by_name, get_entity, search or answer_sources, one daily counter: the tool's name and whether the lookup found something or nothing. On an MCP mount the answer is read from a copy for that alone, and nothing of it is kept. It is a usage counter like the others (usage_counters), and nothing else.
  • What it never holds. The question's text, the name that was looked up, the answer, an IP address, an install id, a session, an account or a key. A miss says that a lookup found nothing, not what was asked.
  • Where it is read. The miss rate per tool, beside the question bank's own gap lists, chooses the next data wave. It is never shown per person, because nothing per person exists.
  • What would need a new decision. Recording which names fail to resolve, even without the question, is a different thing: it keeps game names players typed. It is not built, and it needs its own line here first.

Who is asking, in /ask

Decided 2026-10-29 by Carlos: use it for the answer, and don't collect more than that. A question asked with /ask carries the asking character's class (the game's token, such as MAGE), level and current zone, beside the realm it already carried. They are there so "where is the flight master here" or "best talents for my class" can be answered:

  • They wait in the addon's question queue in the player's own SavedVariables file. Uploads never carry that queue: the Desktop App, the uploader and the upload page send recorded rows only.
  • The Desktop App or companion reads them for that one answer: "here" becomes the zone and "my class" the class in the lookup, and a model the player configured is told who is asking. A lookup sends only the question's words, as before.
  • The local answer store keeps the question as it was looked up ("here" as its zone), for follow-ups, and never the class or level. Nothing on the site keeps any of them; the question ledger counts tools, as above.
  • Any wider use (collecting them, counting them by class or zone) is not built and needs its own line here first.