Straight answers.

A browser extension that hides posts you don't want to see, by meaning rather than keywords, using models that run on your own device.

X, Reddit, LinkedIn, and YouTube at launch, with more added as feeds are mapped. The filter itself is site-agnostic: if content renders in your browser, Sift can judge it.

No. Free works with zero sign-up. An account only exists for Pro subscribers who want sync or billing.

Highlight any text in your feed and choose "Sift this". Sift asks what you meant by it, offers scopes (say, real-world violence vs. fictional), and builds the preference. It takes under a minute.

Keywords match strings; Sift matches meanings. A post saying "a man was brutally attacked outside the stadium" contains no banned word but matches a violence preference. Conversely, "violence" in a film-review context can be allowed while real-world footage isn't.

Sometimes, which is why nothing is hidden silently. Every hidden post carries a one-click receipt: what matched, why, how confident. Plus "Show anyway". Every correction teaches your local model where the line sits for you.

On Free: alt text, captions, OCR, thumbnails metadata: the cheap signals. On Pro: full visual and audio models, sampled intelligently rather than frame-by-frame, so your battery survives.

No. Filtering happens locally after content loads. Nothing is posted, clicked, or signaled back to the platform.

Because for us it's free to run: matching happens on your device against preferences already saved there. Our servers aren't in the loop, so there's no per-scan cost to pass on. Pro pays for genuinely expensive work: visual models, audio transcription, cloud LLM calls.

You stop being able to add an eleventh. That limit is enforced by your account, not by a number in the page. Existing filters keep working fully. You can delete one to make room, or upgrade for unlimited slots.

We deliberately skipped time-limited trials: they expire before the habit forms. Instead, Free gives you the real product capped at ten filters, forever. Upgrade when you outgrow it, not because a clock ran out.

Not on Free. Ever. On Pro, only a specific ambiguous post's text excerpt may go to the cloud LLM, only for preferences where you enabled cloud checks, and responses are discarded. Details on the privacy page.

Sift uses tiered local models and unloads them when idle. We publish size and latency figures only after validating them on the supported device matrix.

The cascade is built so that most posts are decided by rules in about a millisecond, and the embedding model runs in around twenty. Heavy analysis is rare, cached, and never re-run on content you've already seen.

It's gone. Preferences, cache, exceptions and reports all live locally and are removed with the extension. Pro users can delete their encrypted sync blob from settings in one click.