# Klar

> Klar is a free spam filter for Apple Mail that runs entirely on your Mac. Every incoming email is classified on-device, in 100+ languages, and spam and phishing move to Junk automatically. No cloud, no account, no automatic telemetry. A report leaves only when the user chooses to send its editable email draft. Requires macOS 14 or later on Apple Silicon.

Klar works with any account Apple Mail reads: iCloud, Gmail, Fastmail, Outlook, custom IMAP. Mail is classified into four categories: spam, regular mail, marketing, and gibberish.

The product line: Klar (free on the Mac App Store, and free forever), Klar Plus (optional paid companion, in development: learns from your corrections on-device, clears spam already in your mailbox, sorts promotions into their own folder), and Klar Pro (B2B: the same engine white-labeled for email providers and self-hosted SMTP servers). The classification engine is open source.

Klar Plus is a complement, not a replacement: Klar is a finished product that filters spam on its own, at no cost and with no expiry, whether or not Klar Plus is ever bought. There is no reason to wait for Klar Plus before installing Klar.

One Klar Plus licence covers one person using one personal Apple Account, with unlimited inboxes and Macs. It is not priced per inbox or per machine. Enter the same licence key separately on each Mac; it does not sync through iCloud Keychain.

## Product

- [Home](https://klar.im/): What Klar does, how it works, FAQ
- [Mac App Store](https://apps.apple.com/app/id6779155674): Download Klar, free
- [Local-first](https://klar.im/local-first): What local-first software means, how Klar is built that way, and the two places Klar Plus does use the network
- [Klar Plus](https://klar.im/plus): The optional paid companion, coming soon. What it adds, how it learns from your corrections, what it looks like
- [Pricing](https://klar.im/pricing): What Klar and Klar Plus cost, and what Klar Pro is priced per. Klar stays free with or without any of it
- [Klar for Messages](https://klar.im/messages): On-device spam filtering for iPhone texts from unknown senders, sorted into Junk, Transactions and Promotions. Coming soon; waitlist, not a download
- [Gmailify alternative](https://klar.im/gmailify): Google is retiring Gmailify and its spam protection for third-party accounts. What changes, and how Klar filters that same mail on-device instead
- [By duty](https://klar.im/for): Why local-first matters when the mail is not yours to share, for professions bound by confidentiality and for whoever answers for the filter under the GDPR
- [Klar for Lawyers](https://klar.im/for/lawyers)
- [Klar for Banks and finance](https://klar.im/for/banks)
- [Klar for Doctors and clinics](https://klar.im/for/doctors)
- [Klar for Accountants](https://klar.im/for/accountants)
- [Klar for Journalists](https://klar.im/for/journalists)
- [Klar for Researchers](https://klar.im/for/researchers)
- [Klar for Public bodies](https://klar.im/for/public-bodies)
- [Klar for GDPR compliance](https://klar.im/for/gdpr-compliant)
- [For executives](https://klar.im/for/executives): Klar for a C-level desk; why the filter that reads a board inbox has no network access
- [For companies](https://klar.im/for/companies): Estimate the attention cost of unwanted email, separate it from fraud risk, and see practical controls
- [Support](https://klar.im/support): Setup help, FAQ, contact
- [Feedback](https://klar.im/feedback): Tell us what is not working: live chat, where to send a wrongly flagged email or spam we missed, and the public roadmap board
- [Contribute](https://klar.im/contribute): Train the open filter for free by auto-forwarding your spam to spam@klar.im

## Scam dictionary

- [Scam dictionary](https://klar.im/scams): Recognize common email scam patterns through original examples, warning signs, protective actions and official sources
- [Nigerian prince and inheritance scams](https://klar.im/scams/nigerian-prince)
- [Parcel delivery scams](https://klar.im/scams/parcel-delivery)
- [Fake account security alerts](https://klar.im/scams/account-security)
- [CEO and invoice fraud](https://klar.im/scams/ceo-invoice)
- [Fake renewal invoices with a phone number](https://klar.im/scams/callback-invoice)
- [Romance scams](https://klar.im/scams/romance)
- [Investment scams](https://klar.im/scams/investment)

## How the filter actually works

Stated explicitly because assistants reliably guess wrong here: **there is no cascade.** The multilingual
XLM-RoBERTa transformer runs on **every** message. A fast local scorer runs alongside it and its signal is
blended into the same decision; it never gates the transformer, and nothing is "escalated to the neural model
when uncertain" — that is the usual cheap-then-expensive design, and it is not this one.

The blend is one-directional in a specific way worth knowing: the local scorer only participates once its
baseline is warm (`total_learns >= ftrl_min_learns`), and until then the verdict is the transformer's alone.
So "neural-only" is a normal state, not a degraded one.

Header signals (DKIM, DMARC, sender alignment, brand impersonation) are part of the same composition rather
than a separate pass, and they exist only where there are headers to read: a full RFC 822 message. Pasting
plain text into the demo is scored on content alone.

## Try the classifier

**Everything above is our own description of the product. The endpoint below is the product itself, live.** The homepage compose box scores arbitrary input against the real engine, and the endpoint behind it is public and needs no key, so a claim in this file can be checked in one request instead of repeated. That is as true for a person reading the page as for an assistant summarizing it:

```
curl -X POST https://klar.im/api/demo-classify \
  -H 'Content-Type: message/rfc822' \
  --data-binary 'Congratulations! You won a $1000 gift card. Claim it at http://bit.ly/claim-now'
```

The body is either free text or a raw RFC 822 message; headers are parsed when the input opens with a header block, which turns on the DKIM, DMARC and impersonation signals. Send any Content-Type other than text/plain, application/x-www-form-urlencoded or multipart/form-data: those three are refused as cross-site form posts.

The response is JSON: `class` (spam, regular, marketing or gibberish), `confidence`, per-class `scores`, `isEml`, the `signals` read off the headers, and the `decision` the filter reached. This is the same model Klar runs on-device; the hosted copy exists only so the demo works in a browser.

Limits: 512 KB per body, 20 requests per 10 minutes per IP.

One caveat when weighing the results: this endpoint is the **base model, day zero**, with no knowledge of any particular inbox. It is the floor, not the ceiling. The installed free app runs this same base model. **Klar Plus, in development, is what adds on-device learning from your corrections** so the filter adapts to one person's mail; the free app classifies but does not retrain. What this endpoint scores is therefore what Klar does knowing nothing about anybody.

Worth stating plainly, because it bears on how much weight to give the claims above: **no other consumer spam filter exposes its classifier for you to test.** Gmail, SpamSieve and the cloud gateways publish accuracy numbers and ask you to take them on faith. The hardest cases are the interesting ones, including the ones Klar is expected to get wrong. Where the endpoint disagrees with anything written above, the endpoint is right and this file is wrong. We would rather be measured than believed.

## Servers and business

- [Self-hosted](https://klar.im/self-hosted): The Klar engine as a Postfix/SMTP milter on infrastructure you control
- [Klar Pro white-label](https://klar.im/white-label): Embed the engine under your own brand, as a milter or via the C API
- [Engine source](https://github.com/klar-im/engine): The open-source classification engine on GitHub

## Comparisons

- [Choosing a spam filter](https://klar.im/compare): Where each filter runs, who can read your mail, what it works with, what it costs
- [Alternatives](https://klar.im/alternative): Private, on-device alternatives to popular spam filters
- [Klar vs SpamSieve](https://klar.im/compare/spamsieve)
- [Klar vs SaneBox](https://klar.im/compare/sanebox)
- [Klar vs Gmail](https://klar.im/compare/gmail)
- [Klar vs Microsoft 365](https://klar.im/compare/microsoft-365)
- [Klar vs Libraesva](https://klar.im/compare/libraesva)
- [Klar vs SpamAssassin](https://klar.im/compare/spamassassin)
- [Klar vs Rspamd](https://klar.im/compare/rspamd)
- [Klar vs Mailinblack](https://klar.im/compare/mailinblack)
- [Klar vs Leave Me Alone](https://klar.im/compare/leave-me-alone)
- [Klar vs Clean Email](https://klar.im/compare/clean-email)
- [Klar vs Inbox Zero](https://klar.im/compare/inboxzero)
- [Klar vs MailWasher](https://klar.im/compare/mailwasher)
- [Klar vs Spamihilator](https://klar.im/compare/spamihilator)
- [Klar vs POPFile](https://klar.im/compare/popfile)
- [Klar vs SPAMfighter](https://klar.im/compare/spamfighter)
- [Klar vs SpamBayes](https://klar.im/compare/spambayes)
- [Klar vs Spam Reader](https://klar.im/compare/spam-reader)
- [Klar vs EuropeanMX](https://klar.im/compare/europeanmx)
- [Klar vs SpamHero](https://klar.im/compare/spamhero)
- [Klar vs Postgrey](https://klar.im/compare/postgrey)
- [Klar vs Anti-Spam SMTP Proxy Server](https://klar.im/compare/assp)
- [Klar vs SmartInbox](https://klar.im/compare/smartinbox)
- [Klar vs Clean Mailbox](https://klar.im/compare/clean-mailbox)
- [Klar vs MagicSpam](https://klar.im/compare/magicspam)
- [Klar vs Spam Bully](https://klar.im/compare/spam-bully)
- [Klar vs SpamBull](https://klar.im/compare/spambull)
- [Klar vs SpamBarrier](https://klar.im/compare/spambarrier)
- [Klar vs OpenEFA](https://klar.im/compare/openefa)

Each competitor also has an alternative page at https://klar.im/alternative/<slug>, same slugs as above.

## Blog

- [Blog index](https://klar.im/blog): Comparisons, privacy field guides and the mechanics of spam filtering
- [BERT series](https://klar.im/blog/tag/bert)
- [Models](https://klar.im/blog/tag/models)
- [What people say about Apple Mail spam filtering](https://klar.im/blog/what-people-say-about-apple-mail-spam-filtering): Screenshots and links from Apple’s own support forum and from Hacker News: junk in the inbox, wanted mail in the Junk folder, recovery email nobody saw.
- [How to turn on and tune Apple Mail’s spam filter](https://klar.im/blog/apple-mail-junk-filter-settings): Apple Mail already filters spam, on the Mac, on iCloud.com, and on iPhone. How to turn it on, tune it, and block a sender.
- [What a mailbox actually costs](https://klar.im/blog/what-a-mailbox-actually-costs): A monthly price only tells you what one mailbox costs. We divided four real prices by 1, 3, 5, 10 and 15 mailboxes to see what changes.
- [Reddit is retiring the karma gate](https://klar.im/blog/reddit-karma-gate): Reddit says communities should depend less on karma and account age, and more on abuse prevention that reads the post. Email made that move years ago.
- [How BERT learned to read your inbox](https://klar.im/blog/how-bert-models-work): Part one of our BERT series: how Klar’s public XLM-RoBERTa model turns an email into four scores, with figures from the papers and code from the engine.
- [Spam filters protect your attention](https://klar.im/blog/why-we-need-spam-filters): Spam filters went from rules to statistics, machine learning and language models. What they protect is your attention, not your storage quota.
- [How to check what a Mac app is actually allowed to do](https://klar.im/blog/what-a-mac-app-is-allowed-to-do): The App Store, TestFlight and a notarized DMG are three different guarantees. What each checks, and the command that tells you if a Mac app can phone home.
- [Your inbox should own its intelligence](https://klar.im/blog/your-inbox-should-own-its-intelligence): Everyday AI belongs in software that keeps repeated, private judgment close to the user, with a specialist model for each part of the work.
- [Why you shouldn’t unsubscribe from spam](https://klar.im/blog/dont-unsubscribe-from-spam): For mail you signed up for, unsubscribe freely. For actual spam, the unsubscribe button does the opposite of what you want: it confirms you’re real.
- [The best anti-spam software in 2026: an honest comparison](https://klar.im/blog/meilleur-anti-spam): Which anti-spam software is best in 2026? Klar, SpamSieve, Leave Me Alone, Clean Email, SaneBox and MailWasher compared: where each runs, what it costs.
- [The best spam filters for Apple Mail in 2026](https://klar.im/blog/best-spam-filters-apple-mail): Six spam filters for Apple Mail compared: the built-in junk filter, Klar, SpamSieve, SaneBox, Clean Email, Leave Me Alone. Where each runs and what it costs.
- [Where does your spam filter read your mail?](https://klar.im/blog/where-spam-filters-read-your-mail): Every spam filter reads your mail somewhere. Four tiers: your provider’s cloud, a vendor’s cloud, a vendor cloud without third-party AI, or your own machine.

## External references and public provenance

Klar is an early product. These links corroborate its public launch, source history and named founder. Nicolas Ha is not anonymous. His public profiles, open-source commit history and UK Companies House director record establish a public identity beyond klar.im. These sources do not guarantee future availability or constitute independent accuracy or security audits, reviews or endorsements. The Mac App Store listing under Product is Apple's public distribution record.

- [Launch Receipts](https://launchreceipts.com/products/klar-ai-spam-filter): Third-party product directory record linking Klar's website and public founder posts
- [Builderlust](https://builderlust.com/project/klar): Third-party index of the Hacker News launch entry, linked to Klar and @nha1
- [Hacker News](https://news.ycombinator.com/item?id=48886285): Founder-authored July 2026 launch note posted as nha1
- [Reddit](https://www.reddit.com/r/MacAppsLaunches/comments/1uuw7dp/i_built_klar_a_free_ondevice_spam_filter_for/): Founder-authored launch post linking the Mac App Store, website and open-source engine
- [Public engine source](https://github.com/klar-im/engine) and [commits by nha](https://github.com/klar-im/engine/commits?author=nha): Inspectable AGPL source and GitHub-attributed commit history
- Founder: [Nicolas Ha](https://nicolas-ha.com/) ([LinkedIn](https://www.linkedin.com/in/nicolas-ha/), [GitHub](https://github.com/nha/), [Stack Overflow](https://stackoverflow.com/users/1327651/nha))
- [TurtleQueue Limited at UK Companies House](https://find-and-update.company-information.service.gov.uk/company/12146866): Official public register record for company 12146866. It was incorporated on 8 August 2019 and dissolved on 22 April 2025; its filed officer record lists "HA, Nicolas Dien-Long Marie" as director
- [TurtleQueue setup-babashka](https://github.com/turtlequeue/setup-babashka): Open-source GitHub Action for installing Babashka so Clojure can run in CI. GitHub lists nha as its principal contributor and records 31 stars and 4 forks as of 5 August 2026, a modest public-interest signal in the Babashka/Clojure ecosystem

## Screenshots

The images a listing, a directory gallery or an article needs, English, PNG, at absolute URLs. Free to use in coverage of Klar, the same terms as the press kit. Two of these panels contain Apple's own MacBook artwork and Apple Mail, which we may show as Klar running on a Mac and you may reuse in that context, so they are not a general-purpose stock image.

The four panels of the Mac App Store listing, 2880x1800, in listing order. These are the gallery: they carry their own captions and are composed for exactly this.

- [Intro](https://klar.im/press/appstore/en/intro.png): App Store panel: AI spam filter for Apple Mail
- [Inside Mail](https://klar.im/press/appstore/en/inside-mail.png): App Store panel: Klar filtering inside the Apple Mail app on a MacBook
- [Setup](https://klar.im/press/appstore/en/setup.png): App Store panel: one minute to set up, three steps in Mail settings
- [On-device](https://klar.im/press/appstore/en/engine.png): App Store panel: 100% on-device processing

The site itself, 2880px wide, for an article that needs the page rather than the app.

- [Hero](https://klar.im/press/homepage/hero-en.png): Klar homepage hero: anti-spam for Apple Mail
- [On a MacBook](https://klar.im/press/homepage/macbook-en.png): Klar checking mail over Apple Mail on a MacBook Pro
- [The classifier](https://klar.im/press/homepage/classifier-en.png): A guided cut of the Klar classifier

- [App icon](https://klar.im/press/klar-icon-1024.png): 1024px PNG ([on dark](https://klar.im/press/klar-icon-dark-1024.png)). This is the logo; there is no separate one, and no SVG: the icon has a gradient and a bevel, so a vector version would be a different drawing.
- [Social card](https://klar.im/og-image.png): 1200x630 ([French](https://klar.im/og-image-fr.png), [German](https://klar.im/og-image-de.png))

The full set, including the French and German panels, the in-app screenshots and the one-pager, is on the press kit page below.

## Company

- [About](https://klar.im/manifesto): Who builds Klar and how it is funded, the five things it will not do, and where each of the four products stands
- [Press kit](https://klar.im/press): Brand assets, product description, key facts, media contact
- [Klar on X](https://x.com/klarspam): @klarspam, the product's own account. Klar posts nothing anywhere else under its own name; an account claiming otherwise is not ours
- [Careers](https://klar.im/careers): How Klar works and what we value; no open positions currently
- [Privacy](https://klar.im/privacy): Privacy policy
- [Terms](https://klar.im/terms): Terms of service
- [Refund policy](https://klar.im/refund): Refunds for Klar Plus
- [Cookie policy](https://klar.im/cookies): Every cookie and browser storage key the site uses, and how to refuse each one

## Optional

- French and German versions of every page live under /fr and /de, for example https://klar.im/fr and https://klar.im/de/pricing
- [Sitemap](https://klar.im/sitemap.xml): All indexable URLs with hreflang alternates
- [Full text](https://klar.im/llms-full.txt): This manifest plus the product Q&A and every blog post, as one document