jev-skip

Skips YouTube sponsors on videos nobody has labeled yet. Catches 77% of the sponsor seconds SponsorBlock's crowd marked across 23 videos, at 34s of false skips per hour, for $0.0008 a video, with answers 0.9s after the request goes out and slices on the bar within about three seconds of opening the page.

npm install && npm run build   # then load dist/chrome-mv3 unpacked, paste your key

Nine seconds, one take: the extension reads the captions, asks once, paints ten slices and jumps 38 seconds of a Squarespace read it was never told about. The popup on the right is the same run, live: 50 segments, 25.3k tokens, $0.0011, answered in 1546 ms. In the clip the answers come from fixtures/answers/, recorded 2026-09-18 and replayed at their recorded latency, because this machine has no key. Everything else in frame is happening.

Numbers from npm run measure over the fixtures in this repo, answered through a Vercel AI Gateway shim rather than the direct API. Re-measure on api.typesafe.ai before pinning anything to jev-1.13.0.

Why

Every sponsor skipper on the market waits for a stranger to submit the timestamps. SponsorBlock is excellent at that and it is always late: a video uploaded an hour ago has no segments until someone watches it, notices the read, drags two handles and submits. For most videos that never happens at all.

This one reads the caption track of the video you are already watching, sends the transcript to Jev once, and gets back a probability per 30-second segment. The bar is painted before the intro is over, on a video nobody has ever labeled. Slices are tinted by how sure the model is, so a borderline sponsor is a ghost and a confident one is solid, and you can see the doubt instead of guessing at it.

The honest limit lives here rather than only at the bottom: it reads text, not audio. No captions, no opinion, and the extension does nothing at all.

Getting at those captions is the fragile part. The caption URL in YouTube's page data is answered with 200 and an empty body unless it carries a proof-of-origin token, which the player mints in the page world. So the extension watches for the URL the player signs for itself and re-asks for the same one as json3. Measured, not assumed: see docs/browser-ground-truth.md. It works today and it is exactly the kind of thing YouTube changes.

Install

git clone https://github.com/valentynkit/jev-skip && cd jev-skip
npm install && npm run build

Chrome: chrome://extensions, developer mode, load unpacked, pick dist/chrome-mv3. Firefox: npm run build:firefox, then about:debugging, load dist/firefox-mv2/manifest.json. Open the popup and paste your TypeSafe key.

variable	required	purpose
key in the popup	yes	your TypeSafe key, stored in the extension, sent only to the endpoint below
endpoint in the popup	no	defaults to https://api.typesafe.ai, point it at a local shim if you use one
TYPESAFE_API_KEY	no	only for npm run record, which rebuilds the fixture corpus

Privacy and permissions

Hosts: youtube.com to read the page you are on, api.typesafe.ai to ask the question. What leaves your browser: the title, channel and caption text of the video you are watching, under your own key. Nothing else, to nobody else. There is no server of ours, no account, no telemetry, and no shared database. The key lives in extension storage that only the background worker touches. On Chrome the content script is refused access to it, which was checked against the browser rather than the docs (scripts/keycheck.mjs). Firefox has no equivalent control, so there a content script could read it.

One script runs in the page itself, to see the caption URL the player signs. It reads that URL and the video's title; it never sees the key.

How it works

1. A page-world script keeps the caption URL the player signs for itself, and the content script re-asks for it as json3. If the player has not asked for captions within a few seconds, the extension asks it to load a track and puts your previous choice back. Nothing to read means stop here.
2. Cues are deduped, then cut into 30-second windows snapped to sentence ends, capped at 45s.
3. A regex belt marks windows carrying a URL, a coupon-code shape or "use my link". Evidence code found, not a verdict.
4. The background worker sends the transcript once and one question per segment, then Jev answers all of them in a single request: content, sponsor, intro, outro, self_promo, recap, other.
5. Answers paint the seek bar as they arrive, one hue per category with colour strength from the probability, and arm one timer for the next segment over your threshold. Below it, on other, during an ad, or on any error, nothing is skipped.

Development

npm test          # offline, against scripts/fake-jev.ts, no key and no network
npm run measure   # the headline number over fixtures/, offline
npm run record -- --dry-run   # re-select the 30-video corpus from SponsorBlock, write nothing
npm run record -- --corpus    # fetch captions and crowd labels for the selection
npm run record                # judge the corpus once and cache the answers, needs a key

scripts/reliability.mjs counts how often the bar paints over cold loads, and scripts/record-demo.mjs films a take; both drive a browser that already has the extension loaded, over CDP. docs/browser-ground-truth.md explains why it is done that way.

measure writes measure.json next to the four lines: per-video rows, the threshold sweep, the reliability bins, the pessimistic intersection-over-union figure, and the non-English videos it kept out of the headline. thresholds.json turns the recall floor and the false-skip ceiling into exit codes.

Known limits

• No captions, no opinion. It reads text, not audio.
• The caption read depends on the player signing a URL and on the extension seeing it. Measured 2026-09-19 in Chrome 153 over five videos: 29 of 30 cold loads painted, median 2.9s from navigation to the first slice. That is one machine and one browser, not a promise. A load that misses paints nothing at all; reloading fixes it.
• If YouTube changes how the player requests captions, this breaks. That is one upstream decision away, and the repo has no fallback for it.
• Chrome is the only browser this has run in. Firefox MV2 builds and is untested. Arc silently drops the extension's storage writes, so the key cannot be saved there at all.
• Sponsor versus the creator's own plug is fuzzy, and the crowd disagrees with itself there.
• Unsure segments are painted faint and never skipped. It under-skips on purpose.
• Non-English captions are weaker and are reported separately, never in the headline.
• Every published number came through a gateway shim, not the direct API, on 23 videos. No live call has ever been made from the extension itself.

License

MIT.