The safest transfer is the one that never happens.

Klar filters spam inside Apple Mail, on your own Mac. No message text reaches us, and none reaches an AI provider, because the part of the app that reads your inbox cannot open a connection.

A filter that reads your mail elsewhere is a disclosure.

The GDPR counts "disclosure by transmission" as processing (art. 4(2)). The body of an email is personal data about the sender, and about everyone they name in it, none of whom chose your spam filter. Send it to a vendor that filters on your behalf and that vendor is your processor: an art. 28 contract, an entry in the record you keep under art. 30, an assessment of their security under art. 32. If the mail then reaches that vendor, or one of its sub-processors, outside the EU, that is a transfer, and chapter V wants a mechanism and the assessment behind it.

Filters that call a large language model add a second layer. The message text goes out as a prompt to OpenAI, Anthropic or whoever the vendor picked, usually in the United States, often as a sub-processor listed on a page you have to go and find. The EDPB's December 2024 opinion on AI models is about when a trained model can be treated as anonymous. It does not make the text you send to a hosted model at inference time disappear: that is a disclosure to that provider, and whether they retain it or train on it is a contract term, not a property of the system.

GDPR art. 28 on processors, chapter V on transfers outside the EU

There is no recipient to name.

Klar is a Mail extension. It scores each message on your Mac as it arrives, with a classifier inside the app bundle. Routine filtering uploads nothing, so it adds no filtering processor or international transfer to assess. The optional Email report is separate: a user reviews and sends its editable draft, and our privacy policy explains what that report may contain. If your organisation permits reports, account for that deliberate disclosure in your own process. Your mailbox provider remains a separate arrangement with its own contracts.

The part of Klar that reads your mail is an Apple-sandboxed Mail extension shipped without the network entitlement. It is not that we promise not to send your mail anywhere. It is that macOS refuses the connection, because the permission was never requested. Classification happens on your Mac, instantly.

What that means in practice

No network access: The free App Store app and its Mail extension ship without the network entitlement. macOS refuses any connection they try to open. This is a permission the app does not hold, not a policy it promises to follow.
The model ships with the app: The classifier is a multilingual transformer inside the app bundle. Nothing is fetched when a message arrives, so it works the same on a train as it does in the office.
No account: There is nothing to sign up for and no licence server the free app calls. Unless you contact us or choose to send an Email report, we do not know who you are.
The engine is open source: The scoring engine is AGPLv3. Your IT people do not have to take our word for any of this. They can read it, build it, and watch what it does.

Questions your compliance officer will ask

Do you send message text to ChatGPT, Claude or any other AI provider?

No. There is no API key in the app and no request to put one in. Klar's classifier is an XLM-RoBERTa transformer that ships inside the app bundle and runs on your Mac's own silicon. The free App Store app holds no network entitlement, so even if we wanted to call an outside model, macOS would refuse the connection. The engine doing the scoring is AGPLv3, so this is something your engineers can check rather than something we ask you to believe.

So Klar does not use AI?

Klar is AI. It is a multilingual transformer, the same broad family of model as the products you are being asked to sign a DPA for. The GDPR question is not whether a model is involved, it is who holds the message while the model reads it. A hosted model means your mail arrives at a company you have to contract with, document, and place on a map. A local model means the message is read on the machine it was already delivered to, by software that cannot open a connection.

Does filtering send anything off the machine?

No. The free App Store app and its Mail extension have no network permission. If you choose Email report, Klar opens an editable draft in your own mail app; nothing leaves until you send it. Do not include confidential correspondence unless you are authorised to share it. Klar Plus, the paid companion you install yourself, separately uses the network for its licence, model and version checks. None carries your mail. The full list is on the local-first page.

Do we need a data processing agreement with you?

That is a question for your own counsel, and it depends on which part of Klar you run. Routine filtering sends us nothing. A user who sends an Email report may share selected messages, diagnostics and a removable monthly usage summary; our privacy policy explains what we then hold. Decide whether staff may use that report action with confidential correspondence.

What we are not claiming.

We are not offering legal advice, and a spam filter does not make anyone compliant. Your obligations are yours, and your bar, your regulator or your DPO is the right place to test them. What we can tell you is exactly where the software runs and what it is permitted to do, which is the part most vendors leave you to infer.

We are also not claiming perfect filtering. No filter is, and any vendor quoting you a single accuracy number is quoting a number from their corpus, not your inbox. What we will say is that being wrong about a message on your Mac and being wrong about it on someone else's server are not the same kind of wrong.

Read the whole local-first argument

Try it on the inbox you cannot risk.

macOS 14 or later, Apple Silicon. No account, no cloud classification, no setup. It filters inside Apple Mail on your Mac.