Sweep

Sweep your feed clean. In plain words.

Write a rule such as AI replies or posts that brag. Sweep reads each post in your feed and blurs the ones that fail a rule. Each blurred post shows the reason, and one click shows the post again.

It works on X, Reddit, LinkedIn and Hacker News.

What you can do

You write	Sweep does
Hide AI reply	Blurs replies that read as AI-generated
Hide posts that sell a course	Blurs them, with the score on top
Show only high quality	Blurs each post that is not high quality

• Two rule types: Hide and Show only. You can mix them.
• A scope for each rule: all, replies only, or posts only. Jev reads your new rule and chooses the scope, and you can change it with one click. (Sweep sees replies on X post pages. On the other sites, each item is a post.)
• Two results: blur the post (it keeps its place and its size, with the reason) or remove it.
• No jumps: the blur fades in. In remove mode, a post below the screen goes at once, and a post that you see closes softly.
• One slider: from gentle to strict.
• Ready rules: AI reply, slop, engagement bait, crypto, brag, rage bait, only high quality.

Install

1. Download this repository (Code → Download ZIP) and unzip it.
2. Open chrome://extensions and turn on Developer mode.
3. Click Load unpacked and select the folder.
4. The welcome page opens. Paste your TypeSafe API key and click Save.
5. Open a feed.

You need a key for the TypeSafe API. Sweep has no server of its own: you bring your own key.

How it works

Sweep uses Jev, a small model from TypeSafe that does not write text. It answers a yes/no question with a probability, in about one second.

1. A post comes near your screen.
2. The extension asks Jev one question for each rule: "Does the post match this description?"
3. A Hide rule fails the post when the answer is "yes" with more certainty than your slider. A Show only rule fails it when the answer is "no".
4. The post gets a blur and a label, for example AI reply 94%.

All rules for one post go in one request. A post is checked one time for each set of rules. When a check fails, Sweep tries that post one more time, and then it stops.

Privacy and security

What leaves your browser

• The text of a post that comes near your screen, the text of the original post when the post is a reply, the name of the site, and your rules.
• They go to api.typesafe.ai only, with your key.
• Nothing more: no author names, no cookies, no page addresses, no analytics. There is no Sweep server.

How the key and the rules are kept

• The key, the rules and all settings are in chrome.storage.local. The extension sets that area to TRUSTED_CONTEXTS, so content scripts, and the web pages they run in, cannot read or change the store. The worker sends the feed script a copy of the rules and the display settings only: no key, no limit, no counter.
• Nothing syncs to other devices. Only the service worker sends the key, and only to api.typesafe.ai.
• The extension checks a new key with one test request before it saves it.
• The feed page gives the worker the post text only. The worker adds the rules from its own store, so a page cannot send its own questions with your key.
• The labels on a blurred post sit in a closed shadow root. The scripts of the feed page can see that a post is blurred, but they cannot read your rule text.

Other guards

• The only permission is storage. The only host permission is https://api.typesafe.ai/*.
• The service worker accepts messages from this extension only, and filter requests from the five feed sites only.
• Each request has limits: 12 rules, 300 characters for a rule, 1,500 characters for a post.
• A daily limit (5,000 posts, you can change it) counts each post before its request goes out. When the server is busy, one request makes up to 4 attempts. One post uses 2 requests at most for one set of rules, and each request takes its own place under the limit.
• The worker reads the on/off switch before each request. When you turn Sweep off, the waiting requests and the retries stop. The queue holds 80 requests at most.
• No remote code, no eval, no innerHTML with page content. The fonts are in the package.

Limits you must know

• Jev makes errors. Use the blur mode first, and turn on "Show the scores" to tune a rule.
• A post can contain text that tries to trick the model. The worst result is a wrong blur.
• The feed sites change their pages. When a site changes, a selector in content.js can stop working. Open an issue.

Cost

The popup shows an estimate for today. With three rules, one post uses about 500 input tokens (measured). One thousand posts cost about $0.02 at $0.042 for one million tokens. Check the TypeSafe price page for the current price.

Add a site

A site is one entry in ADAPTERS in content.js:

{
    site: 'Reddit',
    host_re: /reddit\.com$/,
    selector: 'shreddit-post',                       // the post elements
    text: el => el.getAttribute('post-title') ?? '', // the text to check
}

Add the site to matches in manifest.json and to FEED_HOSTS in background.js.

Credits

• Judgments: TypeSafe Jev
• Fonts: Bricolage Grotesque and Figtree, SIL Open Font License (see fonts/)
• Licence: MIT

Sweep is an independent project. It is not connected to X, Reddit, LinkedIn, Y Combinator or TypeSafe.