Tidy your Bilibili account from one Chrome extension: sort favourites into the right folders with an AI you choose, and unfollow the accounts that stopped uploading — every change goes through a review table first, and every batch can be undone.
English · 繁體中文
Bilibili gives you favourite folders but no help deciding which video belongs where, so a backlog of hundreds ends up in the default folder for good. It gives you a follow list but no view of who has stopped uploading, so the list only ever grows. biliTidy does the deciding for you and leaves the confirming to you: it reads your data through the login you already have, lays it out as a review table with the evidence on every row, lets you tick and confirm, executes in batch, and can undo. Nothing moves and nobody is unfollowed until you press the button.
| Organise favourites | Clean up follows |
|---|---|
 |
 |
Requires Node.js 20+ and Chrome or Edge 116+.
npm install
npm run build # produces .output/chrome-mv3Then chrome://extensions → Developer mode → Load unpacked →
select .output/chrome-mv3.
Log in to bilibili.com, then click the extension icon. It opens one tab with four pages: Organise, Folders, Follows, Settings.
- Settings → 01 AI endpoint — fill in Base URL, API Key and Model of any
OpenAI-compatible endpoint, press "Test connection". Only
https://addresses are accepted, plushttp://localhostandhttp://127.0.0.1for a local model; anything else is refused rather than quietly swapped for another endpoint. Saving asks for permission to that one origin. (Settings are grouped by who uses them: chapters 01–03 only matter for organising favourites, 04–06 are shared; the follow clean-up never needs an AI endpoint.) - Folders — write one sentence per folder describing what belongs in it. Type it, import the folder's Bilibili description, or have the AI summarise what is already inside; both start as drafts you have to accept.
- Organise — pick a source folder, tick the folders it may move into, run the classification, review the table, then execute. Undo puts everything back in one click.
- On any video page a ✨ Smart favourite button next to the native one files the video into the best folder in about a second, with a 10-second window to undo or pick manually.
Write the descriptions first. In a measured run, changing one folder's description from a bare title like
游戏视频to an actual sentence took the hit rate from 94% to 100%; attaching cover images to every video cost 62 extra requests and did worse. (the evaluation)
- Follows — the first screen tells you how many accounts you follow, how many requests the run will take and roughly how long. Press Read the list and check. One request per account at the read rate in Settings; a couple of thousand follows is half an hour at the default 2 req/s. Results are cached for 30 days, so the next run only asks about new accounts.
- On the review table set the threshold (180 / 365 / 730 days, type one, or drag the playhead on the timeline column — every row draws the account's silence from its latest upload to today), filter by group or kind of follow, tick the rows, press Unfollow N accounts and confirm. Undo follows them again and puts them back in their groups.
Shared by both:
- Review is the product. Every row shows the evidence: the AI's suggestion, reason and the fields it used; or the account's latest upload, how long ago, its groups. Nothing is pre-ticked that the tool is unsure about.
- Unconfirmed never enters a batch. Videos that are no longer available are never sent for classification; accounts whose upload status could not be confirmed (a rate-limited request looks exactly like "never uploaded") are shown but cannot be ticked, unfollowed or exported.
- Rate limiting. All reads go through one throttled queue (default 2 req/s
± 30% jitter; Conservative is 1 req/s), writes are serialised. On
-352/-412/-799/ HTTP 412 the run backs off 60 → 120 → 240 s and then stops, keeping its progress — rows already fetched are reviewable, the rest can be checked later with one click. - One long task at a time. Organising and the follow clean-up share the rate limiter, so while one runs the other page's main button is greyed out with the reason next to it.
- Runs happen in the extension's own tab, can be cancelled, hold a Web Lock so Chrome does not freeze them in the background, and are snapshotted so closing the tab keeps the review. Keep the tab in the foreground when you can — Chrome clamps background timers to 1 s, which roughly halves the read rate.
Favourites:
- Staying put is the default. The prompt starts with where the video already lives; with no clearly better home, it stays. A video can live in several folders, so the model answers "which folders should this end up in" and can suggest keeping it and copying it elsewhere ("Also keep in place").
- Transparent cost. Settings computes which requests your configuration will make; the Organise page recomputes it for the actual number of videos.
- Deleting a folder is not implemented. On Bilibili that destroys the videos inside it and cannot be undone.
- Download the run as a CSV — every row, with what the model suggested next to what you chose, its reason and the fields it used.
Follows:
- Three states, not two. Every account is has videos, verifiably has none, or unknown. Unknown rows stay visible but locked; the next run retries them.
- Groups are a signal. A follow group (or special follow) usually means you filed the account on purpose; filter by group, or "no group" only.
- Unfollowing is one request per account, spaced out, with a second confirmation in the action bar (Bilibili's batch endpoint only supports follow and block). Undo re-follows in batches and restores groups, including special follow; quiet follows come back as normal follows.
- Export is still there as a fallback: copy the ticked UIDs (comma separated, the format the Greasy Fork follow manager imports) or download a CSV. Both exclude unknown accounts.
| Permission | Why |
|---|---|
host_permissions |
api.bilibili.com (API), www.bilibili.com (cookie and the video-page button), *.hdslb.com (covers and subtitles) |
optional_host_permissions |
Requested only for the AI origin you enter, when you save it. Cleartext HTTP is allowed only for localhost |
content_scripts |
Only on www.bilibili.com/video/*, to add the ✨ button. It makes no requests itself |
cookies |
Reads bili_jct for the CSRF token on writes (moves, unfollows, follow-agains) |
declarativeNetRequestWithHostAccess |
Adds Referer/Origin to the extension's own calls to api.bilibili.com — the follow-list endpoint returns an empty list without them |
storage, unlimitedStorage |
Settings, descriptions, run snapshots, the IndexedDB caches |
Your API key is stored in chrome.storage.local, is never written to a log, a
snapshot, or an export, and is only ever sent to the origin you configured.
Everything else stays in this browser; the only Bilibili host it talks to is
api.bilibili.com (plus hdslb.com for cover thumbnails and subtitles).
| Command | |
|---|---|
npm run dev |
WXT dev mode with HMR |
npm run build / npm run zip |
Build / package |
npm run typecheck / npm test / npm run lint |
tsc --noEmit / vitest / oxlint |
npm run ui-preview |
Runs all four pages against mocked Bilibili and AI endpoints — a whole classification, a whole follow clean-up, the mutual-exclusion state, settings, Traditional Chinese — and screenshots every screen. No account, no API key, no network. |
npm run quickfav-preview |
Same trick for the video page: feeds a fake video page to the real content script and screenshots the smart-favourite toasts |
npm run smoke |
Playwright loads the built extension and walks the settings page |
npm run gen-icons |
Regenerates public/icon/*.png from icon.svg (renders the SVG in Playwright's Chromium) |
The last four commands drive Playwright's own Chromium — run npx playwright install chromium once before the first of them.
.env.example is only for manual test scripts. The extension never reads it —
the real Base URL and API key are entered in Settings.
The in-depth docs are written in Traditional Chinese.
| docs/how-it-works.md | What each button does and what every setting changes |
| docs/design.md | Why it is built this way, and the Bilibili API findings behind it |
| docs/research/ | The measurements quoted above |
| PRODUCT.md / DESIGN.md | Who it is for and what must not change / the visual system |
| CHANGELOG.md | What changed |
biliTidy is the merge of two earlier private extensions by the same author, biliFavOrg (favourites) and biliFollowCleaner (follows). They did the same kind of work, so they now share one toolbar icon, one settings page, one rate limiter and one look; this repository is where their combined history continues.