Sieve
A Chrome extension for developers who build with coding agents. It turns techniques worth trying, from
LinkedIn, X, Reddit and YouTube posts, into briefs your coding agent (Claude Code, Codex, Cursor and others) can
try in your repo, and it scores those feeds so you see which posts are worth your attention. It never clicks, comments or posts for you, and never installs
or runs anything: a brief is text you copy.
A paid Sieve account is coming: no API keys to manage, and one library across Chrome and iPhone (Sieve for iPhone is
in App Review). Join the waitlist. The extension stays free and open source.
Scoring runs on Jev, TypeSafe's decision model, reached through your OpenRouter key (or
directly, with a TypeSafe key, if you have one and no OpenRouter key). Sieve is an independent project, not made
by TypeSafe. Jev doesn't generate text: it picks from answers you define and returns a probability. When
OpenRouter can't reach Jev, a general model is asked the same questions through OpenRouter and answers in the
same shape. Either way the scorer makes the narrow calls (is this
worth reading? could I answer this?), plain code handles the rules, and
you do the part that needs judgement.
What it does
LinkedIn feed
- Every post that stays on screen for half a second gets a score. 0.7 and up gets a blue bar and a badge
with the kind of post and the topic. 0.4 to 0.7 gets a light bar. Below 0.4 fades
(hover to read; can be turned off). Promoted posts are skipped. Reshares include the shared post.
- Technique posts show a Brief button (below).
Reddit (www and old.reddit)
- Different question: is this someone asking for help that you could answer from first-hand experience?
The badge also shows the thread's age and comment count, and whether it's still fresh
(≤ 12 hours, ≤ 40 comments).
YouTube
- Every video tile you scroll past gets a score chip, from its title, channel and length and whatever else
says what the video is: on search pages the description lines, chapter titles and summary YouTube shows
with the tile; on home and sidebar tiles, which show only a title, the start of the video's description,
which Sieve asks YouTube for (its own player endpoint, without your YouTube cookies; switch it off in
settings). A joke or clickbait title on a real tutorial no longer sinks it.
- Watch it for me: a video model (Gemini 2.5 Flash-Lite via OpenRouter by default) watches the whole
public video, picture and sound, and returns watch / skim / skip, key points with clickable timestamps,
the best moment, learnings and claims to check. About 0.2 US cents per minute of video, up to 55 minutes.
Every result is saved to the daily learnings.
X
- Same scoring and briefs as LinkedIn. Ads are skipped.
- On a post's own page, Brief reads the author's whole thread from X's own page data (the full text, not the
"Show more" cut), and looks at the pictures in it. An X video gets Watch it for me, as on YouTube.
Technique briefs
- On LinkedIn and X, posts that teach a technique or tool you could try get a "technique to try" badge and,
unless they score low, a Brief button. On YouTube, Watch it for me writes a brief when the video teaches one.
- A brief says what it is, what the author claims (labelled as theirs), what to check, what you need, a 15 to
30 minute way to try it, and whether it's worth keeping as a skill.
- Copy as prompt puts the brief on your clipboard for your coding agent. It starts with a note that it's
third-party material, that the agent should treat it as information and not instructions, and that it should
ask before running any command. When the item has a Sieve id (
chrome:..., the id your agent sees once
Your agent is on and the item has been sent), it's the last line, so an agent with Sieve's tools can find
the brief and, if you agree, record how it went. A brief with a warning gets the id alone, with no offer to record anything.
- When the model notices text aimed at AI agents in the source, or Sieve's own check finds a blatant case or
hidden characters, the brief carries a warning at the top. Sieve's own code then makes "Check the source
before copying anything from it." the first step, drops any step or need with a link, a download, a pipe
into a shell or a package install, marks the brief not worth a skill, and ends the copied prompt with an
instruction to show you the warning first and not fetch, install or run anything from it unless you ask.
This lowers the risk; it doesn't remove it: the model can miss a passage, and the rest of the brief can
still quote the source.
Daily learnings
- Posts scored 0.7+, and any post you brief, are saved locally for 30 days. The digest page summarises them
into what people built, numbers worth remembering, patterns, people worth a conversation, and open questions,
with every bullet attributed to its author. Reddit threads you could still answer are listed on top.
- The digest model gets your saved posts as data, with a rule never to follow text in them aimed at AI tools.
A post where Sieve's own check finds a blatant case or hidden characters isn't sent at all, and the digest
ends with a "Left out" line that says how many and whose. This lowers the risk; it doesn't remove it: the
model can still miss a passage.
- Optional daily reminder notification (18:00 by default), only when something new was saved, not counting
posts the digest leaves out.
- Technique briefs from the last 30 days are listed with Copy as prompt. Export library saves everything
Sieve kept (posts, videos, briefs, digests) as one JSON file you own.
Make it yours
Settings (right-click the icon → Options, or Settings in the popup). Since 1.5.0 the page follows Sieve for
iPhone's design: Your agent comes first, and every change saves as you make it (a field when you leave it,
a switch or choice at once), with a short "Saved" note. Keys save only through Check and save keys:
- What you care about: one line about who you are and up to 8 topics. Every post is scored against these,
and the badge shows which of your topics a post is about.
- Kinds of posts to show on LinkedIn and X (technique to try, built something, opinions, questions, news,
promotion, personal) and on Reddit (asking for help, discussions, showcases, rants, news, promotion); YouTube
has its own list.
- Text aimed at AI tools: a post that talks to whatever scores it (asks to be rated highly, fakes a
scoring field, tells a model what to answer) lands low with the reason "text aimed at AI tools", whichever
scorer read it. Plain code catches the forms that can't be anything else, and when the general model scores
instead of Jev, it also reports any passage aimed at it. The trade-off, chosen on purpose: a post about prompt injection that
quotes a payload without quote marks can land low too.
- Rules that always win over the score: words that always show a post (a person, your company) and words that
never do (crypto, webinar, "we're hiring").
- Scores: where highlighting starts, what counts as low, and whether low posts fade, hide or stay.
- Reddit: only in the subreddits you list, and your own definition of "still fresh".
- LinkedIn, X, Reddit and YouTube can each be switched off from the popup. A saved change re-scores what's
on screen in open tabs.
Setup
chrome://extensions → Developer mode → Load unpacked → this folder.- Add a key. Settings open on install with Start here (Connect OpenRouter, or paste a key), and the popup
shows Add your key until a key is saved. Once a key is saved from Start here, it shows four feed buttons
(LinkedIn, X, Reddit, YouTube) and step 4: when a post gets the "technique to try" badge, press Brief, then
Copy as prompt and paste it into your coding agent. The keys:
- OpenRouter API key (scoring posts with Jev, briefs, digests and Watch it for me). Jev scores through this key, about 3 to 5 US cents per 1,000 posts; briefs and digests use
deepseek/deepseek-v4-flash by default. Checked against the API before it's saved.
Start here's Connect OpenRouter gets one through OpenRouter's sign-in window (OAuth with PKCE): OpenRouter makes a key labelled "Sieve for Chrome" in your account, and it's checked and saved exactly like a pasted one. Paste a key instead goes to the field.
- TypeSafe key (optional), under the OpenRouter key: a TypeSafe API key is only used when no OpenRouter key is saved, and then Jev scores directly at TypeSafe. Briefs, digests and Watch it for me still need an OpenRouter key. The old "Score posts with Jev" switch is gone (since 1.4.1): with an OpenRouter key saved, Jev scores through it whatever the switch was set to. The settings page says which key scores.
- Your facts for Reddit. Only true, first-hand things. The scorer uses them to judge whether you could answer a thread.
- Reload LinkedIn, X, Reddit or YouTube and scroll.
Keys live only in chrome.storage.local in your browser profile. Sieve sends data to two services at most and
nowhere else, unless you turn on sending to your agent (see Your agent below). Scoring requests go to OpenRouter, which passes them to TypeSafe for Jev; when OpenRouter can't
reach Jev, the same request goes through OpenRouter to a fixed general model (DeepSeek V4 Flash) instead; with
only a TypeSafe key, they go straight to TypeSafe. Whoever receives a scoring request gets what is scored: a
post's text (for YouTube, the title, channel and length, the text YouTube shows with the tile, and for a tile
that shows none, the first 1,500 characters of the description), your role and topics, and on Reddit your
Reddit facts. OpenRouter gets a post's text when you ask for a brief (with your role and topics), your saved posts (except any Sieve left out) when you ask for a digest, and a
video's link, title and channel (with your role and topics) when you have it watched. OpenRouter passes each
request to the model you picked, except scoring, which goes to Jev (or, as the fallback, DeepSeek V4 Flash).
Cost (October 2026 prices)
- Scoring with Jev through OpenRouter (the default): about 1,000 input tokens per post at $0.042 per million,
about 3 to 5 US cents per 1,000 posts, paid from your OpenRouter credit. Reading YouTube descriptions (26 Sep)
raises YouTube from about 3.7 to about 4.5 US cents per 1,000 videos (about 1,075 input tokens on average on 20
real videos that all got a description, against 880 from the title alone). The description request itself
is free: about 5 KB from YouTube per bare tile. With only a TypeSafe key, Jev direct has the same token
price, billed by TypeSafe.
- The chat model is only a fallback, for when OpenRouter can't reach Jev: about 6 US cents per 1,000 LinkedIn
and X posts, 9 on Reddit (your Reddit facts go with each post) and 5 on YouTube (at most about 7 with
descriptions), with DeepSeek V4 Flash, measured on the invented sets with
test/compare_scorers.mjs on 24
Sep 2026.
- A digest of a day's posts: well under a cent.
- A brief: $0.00004 to $0.0002 with DeepSeek V4 Flash (measured on invented posts).
Your agent (invite-only)
Off by default, and for now only for people Sieve has invited. Settings > Your agent sends your Chrome library
to Sieve's server, so a coding agent you connect there can read it next to your iPhone pins.
- What is sent: your saved posts with their text, your briefs and your watched-video notes, the same items
as Export library, but without digests. Nothing else leaves your browser for this.
- When: only after you turn it on. It sends at once on Turn on and Send now. After that, within 15
minutes of a change to your saved posts, briefs or watched videos, and only if something changed.
- Where: Sieve's server at mcp.divergada.com, which runs on Cloudflare. You sign in through WorkOS (Google
or email), which gives Sieve your email address. The settings page shows it while sync is on.
- Turning it off: Turn off stops sending, deletes the Chrome copy on the server, signs this Chrome out
and gives back the optional permissions. If the server can't be reached at that moment, the settings page
says the copy may still be there. Delete my Sieve account deletes the account and everything the server
keeps, your iPhone library included.
- Permissions: the
identity permission (for the sign-in window) and access to mcp.divergada.com are
optional. Chrome asks for them only when you click Turn on, and Sieve gives them back when you turn it off.
Connect OpenRouter asks for identity alone, for OpenRouter's sign-in window, and gives it back when
it's done unless sending to your agent is on or signing in.
If usage stats are on (see below), the daily counts also say whether sending to your agent is on, as "on" or
"off". They never carry your email or anything from your library.
Usage stats
Off by default. The popup asks once, and Settings > Usage stats turns it on or off. Consent can only be
changed from Sieve's own pages (the popup or settings), never from a script on a feed page. Only after a yes, once a day, the extension sends Google Analytics
counts for each finished day you used it: posts scored, saved, briefed, copied as a prompt, watched, put
in a digest and exported, per platform, with the extension version, which scorer answered, whether sending to your agent is on, and a random
install ID made in your browser. Never post text, links, names, titles, facts, keys or anything you type.
Turning it off deletes the ID and any unsent counts. All of it is in analytics.js, with a comment at the
top saying exactly this.
It's a plain POST to the GA4 Measurement Protocol from the service worker, so no Google script is loaded
and no permission is needed. The Measurement ID and API secret ship inside the store package, so they
aren't secret; the events are kept narrow so made-up ones can only inflate counts. A build from this repo
has empty values in analytics-config.js and sends nothing. The store package gets the real values from
an untracked analytics-config.local.js: node tools/store-zip.mjs refuses without that file, rebuilds
the config from its two checked lines (so nothing else in the file ships), and --no-analytics leaves
the values out.
Tests
The tests use invented posts (test/sample.json, test/hostile.json, test/reddit_test.mjs). Keys come from
TYPESAFE_API_KEY / OPENROUTER_API_KEY, or on macOS from Keychain items typesafe-api-key / openrouter-api-key.
The live scoring scripts (run_triage, reddit_test, compare_scorers) reach Jev through OpenRouter, as the
extension does, so they need only the OpenRouter key; score_latency also times Jev direct with the TypeSafe key.
node test/rules_test.mjs # the rules on top of the score (offline, no keys)
node test/youtube_text_test.mjs # what a YouTube tile says beyond its title, and asking YouTube for the description (offline)
node test/score_test.mjs # scoring through the real worker: which endpoint answers for each key, Jev's errors, the fallback to the chat scorer, cost, and the chat prompt and parser (offline)
node test/watch_parse_test.mjs # the video answer parser (offline)
node test/brief_parse_test.mjs # technique briefs: shape, safety header, prompt text (offline)
node test/brief_prompt_test.mjs # the post brief prompt and its parser (offline)
node test/linear_regex_test.mjs # the dash and bare-domain patterns: the same matches as before, linear on hostile text (offline)
node test/warned_filters_test.mjs # a warned brief: shells by path, look-alike pipes, missed links, every field rewritten (offline)
node test/digest_prompt_test.mjs # the digest prompt, which posts reach it, the Left out note (offline)
node test/digest_worker_test.mjs # the digest through the real worker: flagged posts never sent (offline)
node test/export_test.mjs # the library export (offline)
node test/brief_storage_test.mjs # briefs and saved posts kept apart per platform, a cross-post digested once, through the real worker (offline)
node test/linkedin_post_id_test.mjs # finding a LinkedIn post's id in the page's data (offline)
node test/watch_drawer_test.mjs # the "watched for you" drawer, built line by line like youtube.js used to (offline)
node test/video_platform_test.mjs # Watch it for me on a platform other than YouTube (offline)
node test/agent_sync_test.mjs # sending the library to your agent: sign-in URL and answers, the body, the hash, the words, and that Copy as prompt's Sieve id matches what is sent (offline)
node test/agent_sync_worker_test.mjs # the worker's sign-in, sends, retries, turning off and deleting, against a fake server (offline)
node test/agent_sync_options_test.mjs # the settings page's Your agent section (offline)
node test/openrouter_connect_test.mjs # Connect OpenRouter: the sign-in URL, the callback, the code-for-key exchange, and Start here's button (offline)
node test/options_autosave_test.mjs # the settings page saves each change as it's made, and on closing the tab (offline)
node test/analytics_test.mjs # opt-in usage stats: counting, consent, the daily send (offline)
node test/analytics_worker_test.mjs # the worker counts actions only after a yes (offline)
node test/store_zip_test.mjs # the store package: leaves out tests and tools, refuses forbidden words and missing loads (offline)
node test/fake_dom_test.mjs # the fake DOM test helper itself (offline)
node test/x_post_data_test.mjs # X's page data: full text, reply chain, pictures and video (offline)
node test/x_thread_test.mjs # X threads: checking page summaries, finding the thread, brief and watch requests (offline)
node test/x_brief_worker_test.mjs # X briefs through the real worker: pictures, the vision model, threads (offline)
node test/x_page_test.mjs # x.js on a fake X page: threads, pictures, Watch it for me (offline)
node test/recycled_tiles_test.mjs # a YouTube tile or X post element reused for another video or post: the old verdict goes, the new one is scored (offline)
node test/theme_contrast_test.mjs # the popup, settings and daily learnings pages' light and dark colours and their contrast (offline)
node test/popup_test.mjs # the popup: Add your key until a key is saved, and the rest of its states (offline)
node test/digest_page_test.mjs # the Daily learnings page (offline)
node test/welcome_test.mjs # a fresh install opens settings, where Start here shows until a key is saved (offline)
node test/brief_test.mjs # briefs for invented and hostile posts, checked (set RUNS=3 to repeat each post, at most 4), under 1 US cent
node test/run_triage.mjs # LinkedIn scoring with Jev through OpenRouter (PREFS=file.json to score as someone else)
node test/reddit_test.mjs # Reddit scoring with Jev through OpenRouter
node test/digest_test.mjs # a digest with three probes aimed at the summariser, checked (RUNS=3 to repeat), under 1 US cent
node test/watch_test.mjs # Watch it for me on one public video (VIDEO=url), about 1 cent
node test/x_live_test.mjs # an X picture brief on Flash-Lite and Flash, and an X video watched, under 1 US cent
node test/compare_scorers.mjs # the invented sets scored by Jev (through OpenRouter) and by OpenRouter chat models side by side, with cost; SET=reddit, youtube or hostile (live, a few cents); TITLE_ONLY=1 with SET=youtube scores the tiles as 1.3.0 did
node test/score_latency.mjs # times Jev direct, Jev through OpenRouter and chat models on five invented Reddit posts, a cent or so
On the invented set, Jev's LinkedIn scoring matched the intended tier on 7 of 8 (a "built a small tool" post
scored high where it was labelled maybe) and Reddit on 5 of 6 (the shared-inbox question scored 0.52, maybe
instead of high).
Limits and fair warnings
- LinkedIn and Reddit markup changes break it. LinkedIn (2026) has no stable class names; the hooks are
[role=listitem][componentkey^="update-card-"] and [data-testid="expandable-text-box"]. Reddit uses
<shreddit-post> attributes. If badges stop appearing, those changed. Reddit shows a note if it can't find posts.
A LinkedIn post's own link comes from the page's React data (linkedin-post-id.js); if briefs start linking
to the author's profile instead of the post, that changed.
- Personal use. It only reads what you scroll past in your own browser and never acts on the page.
LinkedIn and Reddit both restrict automated activity; don't turn this into something that does.
- Scores and briefs are suggestions from cheap models. Check any number before you repeat it.
- If the model returns an empty answer (reasoning models sometimes think until they run out of tokens),
it retries once with reasoning disabled and more room.
Files
prefs.js: default settings, the questions the scorer is asked (built from each user's settings) and the rules applied after.jev.js: Jev, TypeSafe's decision model: which key scores (the OpenRouter key first, a TypeSafe key only without one) and one call to it, with each error and whether to fall back.score-prompt.js: the same questions as one OpenRouter chat prompt, and the answer back in Jev's shape, for the fallback when OpenRouter can't reach Jev.models.js: the default text model for briefs and digests, and the OpenRouter providers Sieve avoids.digest-prompt.js: which saved posts go into a digest (flagged ones left out), the digest prompt, the Left out note and the final text.content.js (LinkedIn), x.js (X), reddit.js (Reddit), youtube.js (YouTube), background.js (API calls, saving, reminder).
linkedin-post-id.js: runs in LinkedIn's own page and finds a feed post's id, so briefs link to the post. Reads only.
x-post-data.js: runs in X's own page and reads a post's full text, reply chain, pictures and video. Reads only.
x-thread.js: checks what x-post-data.js wrote, finds the author's thread, and builds the brief and watch requests.watch-prompt.js: the Watch it for me prompt, cost estimate and parser. json.js: tolerant parsing of model JSON.brief.js: the technique brief (shape, safety header, markdown, Copy as prompt, the rules for warned briefs).
brief-prompt.js: the prompt that briefs one post, and its parser.
brief-panel.js: shows a brief on LinkedIn, X and YouTube. export.js: the Export library file.youtube-text.js: what a YouTube tile says beyond its title (description lines, chapters, YouTube's summary),
and the start of a video's description from YouTube's player endpoint when the tile shows none.agent-sync.js: sending the library to your agent, the pure part (sign-in URL, the library body and its hash, the words). agent-sync-run.js: the worker's part (sign-in, tokens, sends, turning off).openrouter-connect.js: Connect OpenRouter, OpenRouter's sign-in window that gives Sieve a key: the URL, the callback, the exchange, the check, the words.watch-drawer.js: the "watched for you" drawer for Watch it for me, shared by every page that offers it.options.*: settings page. popup.*: toolbar popup. : daily learnings page. All three follow Chrome's light or dark setting: their colours are tokens on , with dark values under .
Building the store package
node tools/store-zip.mjs writes ~/Downloads/sieve-store/sieve-<version>.zip from the current commit
(--ref main for another commit, --out <folder> for another output folder). It won't overwrite an
existing zip unless you pass --force. It packages committed files only, leaving out test/, tools/,
README.md and .gitignore. Before writing anything it checks the package and refuses, naming the
problem, when a packaged file's name or bytes mention a word it must never contain, when it loads a file
that's missing from the package, when a .gitattributes file exists anywhere the commit or the repo could
apply one, or when manifest.json uses a key store-zip doesn't check yet. The word check is a tripwire
against committing the owner's personal copy by mistake, not a guarantee. When the commit has analytics-config.js, the zip needs the untracked
analytics-config.local.js (see Usage stats), or --no-analytics to ship the empty one. A committed
*.local.* file is refused. Commit first, then build.
License
MIT
