What makes it different

Everything Loupe does

The differentiators first, then every feature in each app. Where a number appears, it was measured, and we say where.

Judgments in plain words, run on the device

Write a question the way you'd ask a friend. The Loupe Decision Model (~320M parameters) turns it into a yes/no, pick-one or score decision and runs it on your phone or Mac. It doesn't generate text; it decides.

~10 decisions/s, on-device

Nothing leaves the device

Your photos, files, calendar and contacts are read on the device and never sent to us. No account, no analytics. Airplane mode: everything still works.

0 bytes out

It shows what it's unsure about — and when a keyword search beats it

You only review what it's torn on. Every judgment is measured against the dumb version — a keyword rule — and if the rule wins, Loupe says so and offers to use it.

baseline check on every judgment

Cross-source decisions

One judgment spans everything: “receipts I'll need at tax time” returns email attachments, photographed paper, PDFs and spreadsheet rows in one list.

photos · files · mail · calendar · contacts · web

Watchers

Price rises (+23% insurance at renewal), expiring documents, impersonation (“Mum” from an address or number that isn't hers), fake sites and phishing — each with its evidence.

5 watchers

Verifiable results — “Show where”

Every result names the item, its source and date, with Open original. Privacy findings re-find the match on tap and show where it is, masked.

no black boxes

The privacy check

Finds ID and passport numbers, card numbers, IBANs, keys and tokens, contact lists, payroll sheets and duplicate files in your own sources. Masked previews; nothing proposed happens until you approve.

rules on the device, no model needed

Live run view, with the mascot as the engine

Watch a job run: stages on a loop, question diamonds with their latest answer, one particle per real item, who answered, and the cost of asking a cloud model instead.

EN + AR, mirrored in RTL

Riverflight — watch the model decide live

A river shooter where Loupe flies, with its raw probability bars on screen, against a dumb autopilot on the same seed.

60 fps · p50 ~63 ms

Writing assistant: opt-in, your key, drafts only

Off by default. Bring your own provider key (or Ollama on your network) and it drafts replies and second opinions — labelled drafts that wait for your approval. Judgments never depend on it.

never sends anything as you

Loupe for iPhone

iOS 17+ · English and Arabic

NowThe top finding, what needs you, the privacy check and mail triage cards, and the last sort's real counts (“45 sorted, 32 need you”).
Judgments library55 ready-made templates in 10 categories — search, read the criteria and baseline, “Use this”.
Write your own judgmentPlain words with a live lint: yes/no (options that say what they mean), score with named bands, or pick one.
Results you can checkEach answer with its confidence, source, “answered by rule” when a rule decided, and unsure markers.
Unsure queueOnly what it's torn on, most torn first — plus a random fifth of confident answers, so it can't drift unseen. One tap per answer, Skip, Undo.
MeasureAgreement with you from 10 corrections; calibration (ECE, Brier, reliability) from 30; the model vs the keyword baseline on your corrected items, with “Use the baseline” when it wins.
Threshold sliderMove it and see what would have changed over your logged decisions before you commit.
Five watchersExpiry radar, recurring-money census with a monthly total, term changes, impersonation, site fraud — with evidence, Confirm / Dismiss / Not relevant.
Sort while chargingOff by default. Sweeps every judgment over every enabled source in the background; stops when the phone is hot or in Low Power Mode.
Sources, each off by defaultPhotos (text read on the device), Files and “Send to Loupe” from the share sheet, Calendar (read only), Contacts, Mail (IMAP with an app password, or Gmail read-only — Online).
Items inboxImport CSV/TSV (a row per item), .eml, .mbox, ZIP or pasted text.
Privacy checkIDs and passports, card numbers, IBANs, keys and tokens, contact lists, payroll, duplicate files — masked, with Show where.
Mail triage and link checksPossible phishing and needs-a-reply, with the evidence grouped as warning signs, reassuring facts and “also noticed”.
Review queueAnything a check proposes — remove a duplicate copy, confirm phishing — waits for your Approve.
Preset packsImport a pack of judgments from Files, try the example pack, or export yours.
Web templates (Online, opt-in)Currency, weather and UK trains, with ready questions in English and Arabic or your own; the decision model ranks the fetched rows on the phone, with the rule baseline beside it.
Online phishing checks (opt-in)Known-phishing lists matched on the phone; DNS facts; each result labelled Online with its source.
Writing assistant (opt-in)Your own OpenAI-compatible key or Ollama on your network. Drafts replies and second opinions, labelled, waiting for approval.
Model settingsEvery setting with its value, default and trade-off, and Reset.
Show where / Open originalEvery row names its item, source and date, and opens the original.
Live run viewEvery long job, in place: stages, question diamonds, particles, who answered, cost of asking.
RiverflightFly it yourself or watch Loupe fly, with raw probability bars, a speed panel and a results card against the baseline.
Export my dataYour decisions, corrections and settings as a zip — portable files you own.
The Loupe Decision ModelA consented one-time download (~418 MB, SHA-256 checked). 322M parameters, Apache-2.0 weights, 100+ languages.

Loupe Station for Mac

macOS 14+ · Apple Silicon · powered by the Loupe Decision Model

PlaygroundAsk the decision model typed questions (choice, score, yes/no) about any text or JSON and see every probability. About 40 ms a decision on Apple silicon.
Two checkpoints, routed for youEnglish (ModernBERT-large) and multilingual (mmBERT-base) — Arabic and other languages go to the multilingual one automatically.
Memory modesFull, balanced or low — so the model fits the Mac you have, with unloading when idle.
Workflow templates and template libraryReady-made decision workflows; your business specifics go in preset packs you import.
Folder ScanWhat's in a folder and what looks risky: secrets and tokens, exposed personal or business data, duplicates, large and old files.
Arrangement planA tidier layout proposed with a reason per move; nothing moves until you tick it, and every move can be undone.
Email triageGmail (your own OAuth client), Outlook / Microsoft 365 (your own Azure app) or IMAP. Categories, phishing evidence, needs-reply.
Labels and reply drafts — only with approvalIt may add a label or save a draft in the thread. It never sends, deletes, archives or moves mail.
Decision ledgerEvery decision with its full distribution, so accuracy and thresholds can be measured.
Measure: the decision model against simple rulesEach judgment compared with its keyword baseline on your data.
Calibration, spot-check queue, thresholdsSee where it's over-confident and set thresholds from evidence.
SourcesAccounts, mail archives, scanned folders, statements and attachments.
WatchersThe same five watchers as on iPhone, across your Mac's sources and mail history.
Live run viewEvery long job shows itself in place with the mascot as the engine.
Review queueAgents can propose, never apply. Every proposal waits for a person.
Second opinion (opt-in)A language model you choose re-answers uncertain questions; you decide per question.
Browser protectionAn extension for Chrome, Edge, Brave and Arc scores pages on your Mac — address, forms and page text. Nothing is looked up online unless you turn it on.
Local API with per-agent keysThe server listens only on your Mac (127.0.0.1). Give each agent its own key.
Automatic updatesSigned updates via Sparkle from loupe-ai.com.

Loupe Station — coming soonJoin the waitlist to hear first Coming to theApp Store