---

What it does

• Filters in your own words. "Hide engagement bait and crypto threads", or "only keep software engineering". Whatever you'd tell a person.
• Per-author rules. Never filter someone, or always hide them, straight from a post's ⋯ menu.
• Blocked words, matched before the model is asked.
• Teach it. Mark a hidden post as a wrong call, or pick "Hide posts like this" from the ⋯ menu; your recent corrections go to the model as examples.
• Teach X too. Optionally, Sharp tells X you're not interested in what it hides, so the algorithm learns alongside the model. This is the one feature that acts as you on X; the privacy policy says exactly how.
• Thread escape hatch. One click above the reply box shows every comment in a thread, filters off.
• Nothing shown before it's judged. Undecided posts wait behind a placeholder, so the timeline never jumps. If your provider keeps failing, a post is shown after three tries rather than held back forever: a broken provider should not blank your timeline.
• A classifier by default. A decision model scores every post: fast, nearly free, and it says how sure it is. Close calls stay behind a tinted banner, and you set where the lines fall. Reach it through OpenRouter, Vercel AI Gateway or TypeSafe AI, each with a key of its own.
• Or any chat model. OpenRouter, OpenAI, Anthropic, or your own OpenAI-compatible endpoint, if you would rather have a written reason.
• Images, when they matter. Photos are described in words first, and only for the posts whose text leaves it in doubt.
• Debug mode. Open any judged post to see exactly what was sent, what came back and how long each step took.
• YouTube, the quiet way. Hide Shorts, hide comments, and show, blur or drop thumbnails. Durations stay. No model involved. Works on the mobile site too.
• Greyscale. Wash the colour out of the chrome, the content, or both, on X and on YouTube. Less pull, same information.

Private by construction

There is no Sharp account, no Sharp server, and nothing to sign up for. Your API key lives in your browser and talks to your provider — nobody else sees your timeline, your filters, or what got hidden.

You pay your provider directly. Nothing is collected, measured, or sent anywhere else.

Install

Chrome and other Chromium browsers: from the Chrome Web Store.

Firefox: from Firefox Add-ons. Firefox 140 or newer.

From source, with Bun:

git clone https://github.com/tshmieldev/sharp
cd sharp && bun install
bun run build           # Chrome, into the repository root
bun run build:firefox   # Firefox, into firefox/

In Chrome, open chrome://extensions, turn on Developer mode, click Load unpacked, and pick the repository folder.

In Firefox, open about:debugging → This Firefox → Load Temporary Add-on, and pick firefox/manifest.json. A temporary add-on is gone after a restart; the listing above is the one that stays. Firefox hands out site access one origin at a time, so open Sharp and use Grant access if it says it has none for x.com, youtube.com or your provider; reload any tabs that were already open.

Set it up

1. General › AI — paste a key for the classifier. OpenRouter is the easiest start: one key covers the classifier and image descriptions, topped up a few dollars at a time. A Vercel AI Gateway or TypeSafe AI key works too.
2. X › Filtering — write what you want filtered, then Apply.
3. Optional: X › Advanced sets where a hide starts, how wide the close-call margin is, and which posts get their images described. It is also where you switch to a chat model; set that model up under General › AI, where Browse sorts the catalogue by popularity, price, context or measured latency.

Open X and scroll. ⌥⇧F turns filtering on and off; ⌥⇧T unfilters the thread you're reading.

---

Found a bug? Open an issue. Building on it? See CONTRIBUTING.md.

Privacy policy · MIT licensed · Copyright © 2026 tshmieldev