How it actually works.
The landing page makes the promise; this page shows the receipts. Every number below was measured on the real, shipped extension and is stated with the conditions it was measured under. Where we have no measurement, there is no number.
RECORD CURRENT AS OF 2026-08 · EVERY FIGURE CITES ITS SOURCE AND ITS DATE
01 / THE SHORT VERSION
Three ways to read a tab.
Swoop only ever sees a tab's title and address — never the page itself. It reads them in up to three passes, cheapest first:
The table
4,207 well-known sites are pre-classified inside the extension — sourced from the top ~6,000 sites by global traffic (roughly half of those ranks are CDN and ad-tech infrastructure nobody actually browses, correctly absent), topped up with well-known sites the traffic rankings miss. If your tab is on the table, it's sorted instantly.
The on-device model
Everything else is read by a small language model that ships inside the extension package and runs on your machine — you can see it in the download. On the free tier this is where thinking ends: nothing about your tabs leaves the browser, unless you choose to spend one of the 10 one-time AI credits a free account starts with.
Cloud help, when asked
On paid plans, tabs the on-device model isn't sure about are sent — title and URL only — to a much smarter cloud model, processed, then discarded: never stored by us, never sold, never used to train AI models. How often that actually happens is measured below.
02 / THE SPEED RECORD
One number would be a lie, so here are all three.
On-device sorting speed depends on how many categories the model has to weigh for a given tab. The homepage's "about 42ms … about 180ms" sentence is qualified on purpose — the narrowed case and the widest case are genuinely different jobs. Solid bar: median. Faint bar: 95th percentile.
ON-DEVICE SORT LATENCY · SHIPPED BUNDLE, REAL CHROME · MS PER TAB
Median / p95, measured in the production bundle in real Chromium, over the extension's own messaging path — not a synthetic benchmark (medians rounded from 42.0 / 124.0 / 178.8). Table hits (stage 1) skip this entirely and are effectively instant.
03 / THE ACCURACY RECORD
Measured against people, not against itself.
Whole browsing sessions were labeled by hand, then Swoop sorted them and was scored two ways. Strict: the tab landed in exactly the group the human labeler chose. Lenient: it landed in a group the labeler accepted as right — real tabs often belong to two categories at once, and "Learning vs Development" is a coin the labeler flips too. The truth lives between the ends of each band.
SHARE OF TABS SORTED RIGHT · WHOLE SESSIONS, USER-LABELED · STRICT → LENIENT
Strict = exactly the labeler's group · lenient = a group the labeler accepted. A miss costs one drag — and the drag is remembered, on your device.
One honest limit: the on-device model reads English. A non-English probe scored ~45% — coin-flip territory — so we make no multilingual claim.
nli-eval/results/layer2/session-scores.json · 2026-06 eval campaign04 / WHEN IT ASKS FOR HELP
How often the cloud is actually consulted.
"Cloud help for the trickiest tabs" is a measured share, not a vibe. In the shipped configuration's threshold sweep, the fraction of tabs that escalated past the on-device model was:
0%
FREE — STEADY STATE. THE 10 ONE-TIME CREDITS ARE THE ONLY EXCEPTION, SPENT ONLY BY AN EXPLICIT ORGANIZE.
12.9%
PRO — SHARE OF TABS THAT CONSULTED THE CLOUD
21.5%
PREMIUM — A STRICTER CONFIDENCE BAR, SO IT ESCALATES SOONER
The allowance, precisely.
Paid plans draw from one shared pool — 100 AI actions on Pro, 250 on Premium — covering every AI feature: sorting help, project topics, naming, "Let AI decide". Nothing is unlimited, because a quota-maxed subscriber would cost more than they pay, and we'd rather say so than pretend.
The window is a rolling 24 hours, not a calendar day: it opens with your first AI action and refills exactly 24 hours later. Nothing resets at midnight. The popup shows where your pool stands.
05 / FIREFOX
Same brain. Slower legs.
Firefox runs the identical on-device model with the identical categories — in the 8-configuration browser matrix the scores came out bit-identical. The difference is mechanical: Firefox extension pages can't be cross-origin isolated, so the model runs single-threaded there. Measured at the everyday 12-candidate pass, that's 581 ms per tab on Firefox against 146 ms on Chrome — both browsers on the same machine, same run, so the comparison is apples to apples (absolute numbers differ slightly from act 02's capture, which ran on a different rig). Same answers, roughly 4× slower. Capability parity, not performance parity; sorting still happens in well under a second, moments after a tab loads.
8-arm browser matrix, firefox-tier2-spike · 2026-08-2006 / THE DIALS
Every dial, explained.
Out of the box, Swoop just works — this table is what that sentence is hiding. Every control, what it reaches, and how it ships. Nothing here needs touching, ever; all of it can be.
| DIAL | WHAT IT DOES | OUT OF THE BOX |
|---|---|---|
| Pause | One switch turns all automation off and on. | On |
| Correction reach | Every drag-correction shows a picker for how far the learned rule extends: URL only, URL + subpaths (a Pro width, shown locked on Free), or Domain. | Per correction |
| Manual Overrides | Every learned rule, listed and removable. Changed your mind? Drag the tab back — the newest action always wins (there is no global undo button; each correction card carries its own). Corrections live on your device and never sync between machines; Pro can export and import them as a file. | Empty until you teach it |
| Organize Now scope | The one-click cleanup sorts only ungrouped tabs unless you pick "Re-organize everything". Groups you created by hand are never touched, either way. | Ungrouped only |
| Custom categories | Up to 8 of your own — name, emoji, color — routed to from then on. On Pro and Premium, or free forever after 3 qualified invites. They follow your account to any machine. | None |
| Presets | Remove any of the 12 ready-made categories you never use — all except 📚 Reference, which stays as the fallback bucket. | All 12 active |
| Topic sensitivity | Form a project group at 3 or 4 clearly related tabs (an aggressive 2-tab setting exists in the UI, but the server may still require 3). Ceilings either way: at most 10 new topic groups per run, and pools of 500+ tabs are declined. | 3 tabs |
| "Let AI decide" | Premium only: instead of the fixed categories, the AI drafts a category scheme for the tabs in scope — and every idea waits in a Keep/Discard card. Nothing is ever adopted silently. | Off — your categories |
| Strip tidying | The group you're using stays open while the others fold into compact pills; switch tabs and the strip re-choreographs. Both behaviors are switchable. | On |
| Memory saver | Opt-in and off by default: one switch puts tabs in collapsed groups to sleep to free RAM — they stay visible and wake when clicked. The popup reports the real measured savings when a clean reading is available (always an estimate on Firefox). Never automatic. | Off |
| Duplicates | When the same page is open twice, a "Close duplicates" button keeps the copy you used most recently — a button you click, never a background sweep. The toolbar badge counts what's waiting (correction cards take priority on it). | Badge only |
Everything above lives in the popup or the settings page. What you won't find anywhere: session saving, search, bookmarks or history integration, or anything that closes a tab by itself.
07 / THE PERMISSION DIALOG
About the scary sentence Chrome shows you.
When you click install, Chrome will show a prompt worded roughly like this:
Add "Swoop"?
It can:
Read your browsing history
That sentence is Chrome's blanket wording for the tabs permission, and it claims far more than it grants. What it actually means for Swoop: it can see the titles and addresses of your open tabs — which is the entire input to everything on this page.
What Swoop does not request: the history permission (it cannot read where you've been), host permissions (it cannot read what's on any page — no content, no forms, no keystrokes), and no <all_urls> access of any kind.
| PERMISSION | WHAT IT'S FOR |
|---|---|
| tabs | See open tabs' titles and addresses — the whole input. This is the one Chrome describes as "read your browsing history". |
| tabGroups | Create, label, and color the groups you see. |
| storage | Keep your corrections and settings on your device. |
| offscreen | A hidden extension page where the on-device model runs. |
| contextMenus | The right-click "Swoop → Move this tab to…" menu. |
| alarms | The extension's own internal timers. |
| system.memory | Real before/after readings so the memory saver can report measured savings instead of invented ones. (This and offscreen are Chrome-only; Firefox runs the model in its background page instead.) |
Don't take our word for it: after installing, open chrome://extensions → Swoop → Details and read the permission list yourself. It matches this table.
08 / THE HONESTY LOCK
A test suite pointed at our own copy.
Every load-bearing claim on this site is pinned by an automated test suite: required sentences must render verbatim, and a list of banned claims — inflated counts, unqualified speed figures, invented percentages — fails the run if any of them reappears. We run it against every copy change, including the page you are reading; when a claim drifts from the measured record, the suite goes red until the words are fixed. The figures above, with where they come from:
| FIGURE | WHAT IT IS | MEASURED |
|---|---|---|
| 42 / 124 / 180 ms | Median on-device sort latency at 4 / 12 / 16 candidates (p95 62 / 185 / 254), shipped bundle, real Chromium | 2026-06-11 |
| ~77→90% · ~86→97% | Strict → lenient accuracy on user-labeled whole sessions, free · paid | 2026-06 |
| 4,207 | Pre-classified sites in the built-in table, sourced from the top ~6,000 by global traffic | 2026-08 |
| 0 / 12.9 / 21.5% | Share of tabs consulting the cloud — Free (steady state) / Pro / Premium (P2-016 sweep, 7 sessions / 186 tabs) | 2026-08 |
| 581 vs 146 ms | Firefox vs Chrome per-tab median at 12 candidates, 8-arm matrix | 2026-08-20 |
The house rule behind all of it: a number is measured with its conditions stated, or it doesn't appear.