uFeed

Name the topics you want. uFeed blurs the rest of your feed.

A browser extension for X, LinkedIn and Reddit, on Chrome and Firefox. The model runs on your device. No account, no server, no cloud AI, and no post text ever leaves your browser.

On Chrome, the default multilingual model uses one shared offscreen engine across feed tabs; Firefox and the optional English model keep the per-tab iframe engine. Both paths score on-device and keep page access in the content script.

The idea

Most feed filters ask what you want to get rid of. That list never ends: there is always one more thing to block.

Choose what you want to keep, block, or both. To keep posts about subjects you care about, add topics one per line:

software engineering, programming
machine learning research
video games, game design

Posts about your topics stay as you scroll; other posts are blurred, not deleted. You can also add words or phrases to a keyword blacklist, which blurs matching posts. The blacklist works by itself or alongside topics. With only a blacklist, posts without a match stay visible. One click (or Enter) reveals a blurred post.

Why uFeed

• Choose what to filter. Use topics to keep posts about subjects you care about, a keyword blacklist to blur posts containing words or phrases you choose, or both.
• Your feed stays on your device. An on-device model (197 MB by default) runs inside your browser. Feed text and settings are never sent. On a first install in Chrome or Firefox, the browser opens a welcome page on ufeed.es; the normal page request reveals your IP address and language path to the site host. The model is downloaded once from Hugging Face, and the source is here so you can check how it works.
• Every language by default. Gemma reads every language and is fast enough for a good first-run experience. The smaller English-only model is available in the popup for a one-time 33 MB download. Your topics and settings carry over, and each model keeps its own ratings.
• In your language. The popup and the labels follow your browser's language, and the flag next to the title switches both to another: English, Spanish, German, French, Brazilian Portuguese, Japanese and Chinese (Simplified and Traditional).
• You can see why, and steer it. A 0–10 strictness slider, an option to show every post's score, a peek strip that keeps close calls readable, and ratings that correct near-copies of posts you rated.
• You stay in charge. A blurred post is always one click, or one Enter, away. If anything breaks, the feed shows instead of staying blurred.

What it cannot do

• It reads words, not pictures. A photo with no caption cannot be judged on content. Gemma reads every language; the optional English-only model does not.
• It matches subjects, not quality. A great post and a poor one about the same thing both stay.
• It is not perfect. Stricter settings blur more, including some posts you would have wanted.

The trade-off

uFeed compares the meaning of a post with your topics. It does not reason about the post. A large model in the cloud can judge things like sarcasm or "AI news but not hype" better than a model small enough for your laptop can. uFeed gives up some of that in exchange for privacy, speed, and a filter that behaves the same way every time, so a wrong call can be explained and corrected. How it works has the reasoning.

Install

Chrome — Chrome Web Store.

Firefox — build from source for now (below).

Then open the toolbar popup, add a topic, and visit x.com, linkedin.com or reddit.com. The model downloads once (~197 MB by default) and is cached by the browser. Nothing blurs until it has loaded. If you prefer the smaller English-only model, choose it under Model in the popup instead (a one-time ~33 MB download).

Build from source

Node 22.12+ and npm. Nothing else.

npm install
npm run build            # -> .output/chrome-mv3 and .output/firefox-mv3

Chrome — chrome://extensions, turn on Developer mode, Load unpacked, pick .output/chrome-mv3.

Firefox — about:debugging#/runtime/this-firefox, Load Temporary Add-on, pick .output/firefox-mv3/manifest.json.

Working on it? Use npm run watch instead of npm run dev, and see docs/development.md for why.

npm run check            # format, typecheck, tests, both builds

Tests never touch the network or a live feed. Site adapters run against captured fixture HTML, so a vendor DOM change fails as a red test instead of a silent no-op in production.

Beautifully architected, easy to extend

┌─ entrypoints/ ─────────────────────────────────────────────────────────────┐
│  wires: content · popup · background · iframe · offscreen engine           │
│  ┌─ adapters/ · platform/ ──────────────────────────────────────────────┐  │
│  │  X · LinkedIn · Reddit       storage · messaging · model runtime     │  │
│  │  ┌─ feed/ ────────────────────────────────────────────────────────┐  │  │
│  │  │  the page: find posts, score them, blur, reveal                │  │  │
│  │  │  ┌─ core/ ──────────────────────────────────────────────────┐  │  │  │
│  │  │  │  the rules: scoring · blur policy · settings             │  │  │  │
│  │  │  └──────────────────────────────────────────────────────────┘  │  │  │
│  │  └────────────────────────────────────────────────────────────────┘  │  │
│  └──────────────────────────────────────────────────────────────────────┘  │
└────────────────────────────────────────────────────────────────────────────┘
imports point inward only

uFeed follows the Dependency Rule from Clean Architecture: imports only point inward. The rules in core/ know nothing of the browser, and the page logic in feed/ knows nothing of the extension, so a new site never touches the filter. npm run check fails any import that points outward; wiki-llm/layers.md says where new code goes.

Execution stays separate from these code layers: the content script reads and updates the feed DOM; the iframe runs the per-tab engine; Chrome's offscreen page runs one shared Gemma engine. The background creates that page and connects tabs; it does not score posts. Firefox and Chrome's optional English model use the iframe path.

Read more

• docs/how-it-works.md — where it runs, the model, and every control, in plain language.
• docs/development.md — running, debugging and testing.
• docs/android.md — Firefox for Android.
• docs/privacy-policy.md — what it reads, stores and sends.
• Found a bug or have an idea? Open an issue. Want to change the code? Read CONTRIBUTING.md. Security or privacy problem? SECURITY.md says how to report it privately.
• wiki-llm/ — the authoritative model of the code. Start at the index. Agent rules are in AGENTS.md.

License

Copyright © 2026 Dephelion - Julio Cesar Martin. GPL-3.0.

Use it, change it, share it. If you distribute a modified version, you must publish its source under the same license.

The models are not part of this repository and carry their own terms. Your browser downloads the one you pick from Hugging Face: e5-small-v2 is MIT-licensed, and EmbeddingGemma is under Google's Gemma Terms of Use.