Less noise. More signal.
Know before you click.

winnow-seven.vercel.app · Download · Docs

Winnow — know before you click

Winnow is a Chrome extension that tells you whether an article or video is worth your time before you spend it. It reads the page (or the feed you are scrolling), asks Jev from TypeSafe a fixed set of typed questions, and turns the answers into one of four verdicts: read now, skim, save, or skip. Jev returns probabilities, never prose, so every word on the card is a template filled from typed answers. You bring your own Jev key; nothing goes through a server of ours.

What it does

Page mode. On any article or YouTube watch page, a card in the top-right corner shows the verdict with its confidence, plus insight density (1–10), content type, how likely you already know it, whether its claims are supported, whether it is an undisclosed sales pitch, whether it reads as AI-written, and where the payload sits. On videos with captions, "Skip to" chips seek the player to the segment Jev picked as the core idea. Two buttons, Read and Skip, feed your reader state.

Feed mode. On Hacker News, YouTube home/subscriptions/search, and any page with many external links, each link that scrolls into view gets a small badge (GO, ~, SAVE, SKIP, or ?). Hover or focus it to see the full card. Items are judged in batches of up to 12 per Jev call and cached for seven days, so a second visit to the same front page costs nothing.

Reader state. Your reading goals (free text in options), the topics of pages you have had judged in the last 30 days, and your last few Read/Skip clicks are summarised into every request. "Worth it" means worth it for you. Changing your goals invalidates the cache.

Why typed answers, never prose

Jev is a "System One" model: it answers questions of three fixed types (noul for yes/no probability, choice over named options, score over ordered levels) and returns calibrated probabilities for each. It does not generate text. Winnow leans on that:

• The card can only say things the code has a template for. A model cannot hallucinate a summary, quote, or reason onto your screen.
• Thresholds are numbers you can move in options, and the verdict is a pure function you can read in src/jev/verdict.ts.
• Every judgment can be replayed offline against 40 fixtures with pnpm eval.

The observed API behaviour is written down in docs/jev-contract.md; the code is built to exactly that.

Install

Winnow is not on the Chrome Web Store yet. Install from a release zip:

1. Download Winnow-vX.Y.Z.zip from the latest release and unzip it.
2. Open chrome://extensions, turn on Developer mode, click Load unpacked, and pick the unzipped folder.
3. The onboarding page opens. Paste a Jev API key from console.typesafe.ai and click Verify. Winnow makes one call to GET /v1/models and stores the key only if that succeeds.
4. Optional: click the toolbar icon to open options and type your reading goals.

Cost. Jev bills $0.042 per million input tokens; output is free. One Hacker News front page (30 links) is about 34,000 tokens, roughly $0.0015. A long article is 2,000–8,000 tokens.

Long-form instructions and troubleshooting: docs/install.md.

How it works

1. Extract. The content script pulls readable text with Mozilla Readability (articles) or the caption track via YouTube's InnerTube player endpoint (videos). Feed adapters collect title + snippet for each visible link.
2. Message. The content script sends JUDGE_PAGE or JUDGE_FEED to the service worker. Content scripts never see your key.
3. Judge. The service worker checks the cache, builds one Jev request (state + up to 11 typed questions, or 7 per item in feed batches) and calls POST /v1/systemone. This is the only place in the extension that talks to Jev.
4. Decide. computeVerdict applies ordered threshold rules to the parsed answers and produces a label, a confidence, and templated reasons.
5. Render. The card or badge is rendered inside a shadow root so the host page's CSS cannot touch it and vice versa.

flowchart LR
  P[Page / feed DOM] -->|Readability, InnerTube, adapters| CS[Content script]
  CS -->|JUDGE_PAGE / JUDGE_FEED| SW[Service worker]
  SW -->|cache miss| J[api.typesafe.ai<br>POST /v1/systemone]
  J -->|typed answers| SW
  SW -->|computeVerdict → CardModel| CS
  CS -->|shadow DOM| UI[Card or badge]

Deeper: docs/architecture.md, docs/how-judgments-work.md.

Configuration

Click the toolbar icon (or right-click → Options). The options page has:

Section	What you can change
Reading goals	Free text sent with every judgment. Saving bumps the reader-state version and invalidates the cache.
Modes	Page mode on/off; feed mode on/off; per-site toggles for Hacker News, YouTube, other link lists.
Prefetch link text	Fetch each feed link's page in the service worker so Jev sees body text, not just the title. Requests the <all_urls> permission when you turn it on.
Excluded hosts	One hostname per line; matches the host and its subdomains. http: pages, localhost, and private networks are always excluded.
Thresholds	Eleven sliders, one per rule knob, with a reset button.
Data	Clear the judgment cache; reset reader state (goals, topics, feedback).
API key	Verify and replace the stored key.

Every slider, constant, and default is listed in docs/configuration.md.

Privacy

• Page text, transcripts, feed titles and snippets, your goals text, and the titles of items you marked Read/Skip go to api.typesafe.ai and nowhere else.
• Your key, settings, reader state, and cache live in chrome.storage.local on your machine.
• No analytics, no telemetry, no server of ours.

Full policy, permissions, and how to delete everything: docs/privacy.md.

Development

Node 20 or newer and pnpm 10.

pnpm install
pnpm build        # → dist/, load it unpacked

Command	What it does
pnpm dev	Vite dev server; open /dev/preview.html to render cards and badges from fixture data, no Jev needed
pnpm build	Build the extension into dist/
pnpm typecheck	tsc --noEmit
pnpm test	124 unit tests (Vitest)
pnpm eval	Judge the 40 fixtures with real Jev calls, print confusion matrices, suggest thresholds
pnpm probe	One raw Jev call; prints the response body

src/types.ts          frozen contracts and message protocol
src/config.ts         model pin, limits, thresholds, storage keys
src/background/       service worker: router, storage, the only Jev caller
src/jev/              client, questions, verdict, cache, reader state
src/extract/          Readability, YouTube captions, feed adapters, prefetch
src/ui/               shadow-DOM card and badge, templates, styles
src/content/          page.ts and feed.ts glue
src/onboarding/       key entry on first install
src/options/          goals, toggles, sliders, cache and key management
scripts/eval/         eval harness; fixtures/ holds items and goldens
dev/                  preview page for the UI
docs/                 everything below

Doc	Read it when
docs/architecture.md	You want the module map, message protocol, cache keying, and sequence diagrams
docs/how-judgments-work.md	You want every question and every verdict rule, in order
docs/configuration.md	You want every threshold and constant with its default
docs/editing-questions.md	You are changing what Winnow asks Jev
docs/adding-a-feed-adapter.md	You want badges on another site
docs/eval.md	You are touching questions or thresholds and need to measure it
docs/development.md	Setup, testing conventions, release process
docs/install.md	Install, troubleshoot, update, uninstall
docs/faq.md	Short answers to the usual questions
docs/privacy.md	What leaves your machine and what stays
docs/jev-contract.md	Observed Jev API behaviour; the source of truth

Contributing

Small PRs, pnpm typecheck && pnpm test green, one module per PR. See CONTRIBUTING.md.

Got a wrong verdict? Open a Wrong verdict issue with the URL, the verdict you saw, and the one you expected. Good reports become eval fixtures: a JSON file in fixtures/items/ plus a golden entry, so the mistake is measured every time thresholds move. The recipe is in docs/eval.md.

Known limits

• Paywalls and login walls. The card judges whatever Readability can see. A teaser paragraph is judged as the article.
• Feed items are judged on title + snippet by default. Body-dependent rules (low density, AI-written, unsupported claims, and the read-now density gate) only fire at full depth. Badges lean on content type, goal fit, already-known, sales pitch, and Jev's own verdict.
• Videos without captions are judged on title and description at snippet depth, with no Skip-to chips. Caption URLs embedded in the watch page are token-gated and return empty bodies; Winnow fetches tracks through the InnerTube player endpoint instead, which can break when YouTube changes it.
• Insight density saturates. Polished opinion pieces and tutorials often score 9–10, like original research. Rage bait and advertorials that argue for their own classification can move the answer.
• Accuracy. Against the unreviewed goldens, the held-out split scores 80% verdict agreement and 90% content-type agreement.
• No client-side throttle. The client retries on 429 with backoff but does not pace itself against the account limits (250k tokens/s, 1,200 req/min).

Roadmap

1. Split the insight-density question so "polished" and "original" stop scoring the same.
2. Human-reviewed goldens (reviewed: true in fixtures/golden.json).
3. A caption-path fixture for the YouTube extractor, and a visible failure state on the card when captions could not be fetched.

Questions people ask

What does Winnow actually do? Before you open a link, it tells you whether the page is worth your time: read now, skim, save for later, or skip, with a confidence. On feeds it puts a small badge next to every link in view.

Is it free? The extension is free and MIT licensed. It calls the Jev API with your own key, which costs about a tenth of a cent per Hacker News front page and nothing on a reload, because verdicts are cached for seven days.

Does it send my browsing anywhere? Page text goes to api.typesafe.ai and nowhere else. There is no Winnow server, no analytics, and no telemetry. Your key and your reader profile stay in chrome.storage.local. See the privacy note.

How is this different from an AI summariser? A summary still costs you the read. Winnow answers a decision instead, and it never shows you text a model wrote: every line on the card is a template filled with typed answers. See how judgments work.

Does it work on Firefox, Edge, Brave or Arc? Any Chromium browser with Manifest V3 support should run it; only Chrome is tested. Firefox is not supported.

Can I change what it asks, or how strict it is? Yes. Questions live in src/jev/questions.ts, the verdict rules in src/jev/verdict.ts, and every threshold is a slider in the options page, defined in src/config.ts.

How accurate is it? On a 40-item fixture set with a held-out split, verdicts agree with the goldens 80% of the time and content type 90%. Jev 1.13 is early access. The goldens are still unreviewed. See the eval harness.

Brand

Brand assets live in assets/brand/:

• Mark: mark.svg (Winnow Blue), mark-white.svg, mark-ink.svg; favicon.svg is the blue mark.
• Lockups: logo.svg (mark + wordmark, Ink) and logo-white.svg.
• App icons: app-icon-primary.png (white mark on blue), app-icon-light.png, app-icon-dark.png at 512 px; the extension icons in public/icons/ (16, 32, 48, 128) are the primary icon.
• Social pack in social/$: 1080 \times 1080, 1080 \times 1350, 1080 \times 1920, 1920 \times 1080, 1200 \times 630 ($og.png is a copy), 1500×500, 1584×396, 1640×624, 2560×1440.

Usage: the mark is the rounded three-piece W, as drawn, with no outlines, shadows, strokes, or 3D effects. Keep clear space of half the centre wedge's height around it, and do not go below 16 px (toolbar), 20 px (UI), or 24 px (navigation). Palette, type, badge tokens, and copy rules are in docs/brand/DESIGN.md; a one-page summary is assets/brand/README.md.

License

MIT © 2026 ThinkyMiner.

Credits

Winnow stands on other people's work. Every dependency, typeface, image and source is listed with its licence and link in CREDITS.md. The short version:

• TypeSafe for Jev and its typed-answer API. Get a key at console.typesafe.ai.
• Mozilla Readability for article extraction.
• Vite and crxjs for the MV3 build, Vitest for the tests.
• Cabinet Grotesk, Inter and JetBrains Mono for the type.
• Site imagery generated for this project; the archived draft in site-oasis/ uses Creative Commons photographs credited here.

Winnow is not affiliated with TypeSafe AI.