Skip to content

bloomScroll v0.1.0: first early release

Pre-release
Pre-release

Choose a tag to compare

@rohan2010 rohan2010 released this 18 Oct 10:36
· 7 commits to main since this release

Keeps YouTube on what you came for. bloomScroll is a free, open-source browser extension that removes off-topic videos from your YouTube feed. The tutorials you wanted stay; the rabbit holes disappear. Everything runs on your own computer: no account, no tracking, nothing sent anywhere.

How bloomScroll works

Highlights

  • Filters your feed by intent. Each video on the home page, in search results and in the "up next" sidebar is checked against what you came for. Off-topic videos are removed with a short pixel-dissolve effect, and the gap closes up.
  • Works without any server. A small AI model (MiniLM, 23 MB) ships inside the extension and decides most videos in milliseconds. In tests it removed 16 of 20 off-topic videos and never hid a wanted one.
  • Optional local AI for the hard cases. Run Laya on your computer and it settles the videos the built-in model isn't sure about. It listens only on 127.0.0.1.
  • Four focus modes: Work (tech only), Learn (tech plus your own topics), Relax (everything except blocked channels) and Off.
  • You stay in control. The "N hidden" drawer lists every removed video and why, with one click to bring it back. Always-show and never-show channel lists win over everything else.
  • YouTube clean-up switches: hide Shorts, comments and end-of-video suggestions.
  • Debug overlay for the curious: time until the page settled, what decided each video, per-video timings, and a one-click JSON report for bug reports.
  • One-command setup: bash setup.sh --no-server for the easy path, or bash setup.sh to add the local AI.

Performance

Measured on a 2017 Intel MacBook Pro (2-core i5) with a 40-video test feed:

  • 35 of 40 videos decided on the device, without the local AI
  • Feed settled in about 1–2 s in Chrome (the first page of a session adds 3–6 s while the model loads)
  • Laya, when used, takes about 0.4 s per video, and only sees the unclear few
  • A speed audit cut Laya's cost per video by 2.5× with no loss in accuracy

Install

  1. Install Node.js (LTS).
  2. Download this release (Source code ZIP) and unpack it.
  3. In a terminal, inside the folder: bash setup.sh --no-server
  4. In Chrome, Edge or Brave, open chrome://extensions, turn on Developer mode, click Load unpacked and choose extension/dist.
  5. Open YouTube.

To add the local AI, run bash setup.sh, then server/start.sh, then Settings → Save → Allow. Full steps for newcomers are in the README.

Needs: Chrome, Edge or Brave and Node.js 20+. For the local AI you also need Python 3.10–3.12 (3.11 on Intel Macs) and about 2 GB of disk space.

Privacy

No accounts, analytics or remote logging. The built-in model and Laya both run locally. The only optional network use is Jev, a hosted model you turn on with your own TypeSafe key, which sends video titles and channel names to TypeSafe. The Settings page labels it clearly.

Known limitations

  • Not yet in the Chrome Web Store, so it installs as an unpacked extension (Developer mode).
  • Only tested on macOS; Linux and Windows (Git Bash or WSL) setup is untested.
  • Filters home, search and the watch-page sidebar. Channel pages and playlists are left alone on purpose.
  • YouTube changes its layout often. If the popup says "No video tiles found", please open an issue.
  • Corrections you make are saved but not learned from yet. Calibration is planned.
  • Jev's prompt hasn't been re-tuned since the latest prompt change.
  • After updating the extension, reload any open YouTube tabs.

What's next

👎 on videos that should have been hidden, a calibration button that learns from your corrections, focus-mode schedules, Jev as an opt-in fallback, and a Chrome Web Store listing.


MIT licensed. Built on Laya, Transformers.js and MiniLM.
Found a problem? Turn on the debug overlay, press Copy, and paste it into an issue.