"On-device AI" has become a label that means different things in different apps: sometimes a local model, sometimes a local cache in front of a cloud service. If you are handing an email archive to a tool, the difference matters. This post explains exactly which engines mailin uses, what runs where, how each output is labeled, and what the optional Cloud AI mode does and does not send.
On-device AI email analysis: two engines, no API key
mailin uses two AI engines, both provided by Apple and both running on your device. Neither needs an API key, an account or a network connection, and both are free to use.
The first is Apple Intelligence, using Apple's on-device Foundation Models. This is the engine behind summaries, question-answering and reasoning: the AI Assistant, the narrative in Thread Story, and the AI Digest. It requires macOS 26 or later, or iOS 26 or later, on a device that supports Apple Intelligence.
The second is Apple NaturalLanguage, which handles sentiment analysis, language detection and named-entity recognition. It works on every device mailin supports, back to macOS 14.6 and iOS 17.6.
What happens when Apple Intelligence is not available
If your Mac or iPhone cannot run Apple Intelligence, whether because of the OS version or the hardware, mailin falls back to NaturalLanguage. Sentiment scores, detected languages and extracted entities keep working exactly as before. What you lose is generation: the assistant's written answers and the AI narrative depend on a model that can produce text, and NaturalLanguage is not that kind of model.
The fallback is graceful in the sense that nothing breaks, but it is not hidden. Every AI output carries a provenance tag, described below, so you can always see which engine produced it.
What runs on the device
All of mailin's AI features run locally, on one of the two engines:
- Sentiment analysis on every email, plus topic and keyword extraction and priority scoring.
- A smart auto-tagger that labels messages by topic, sender intent and priority, with manual override.
- The AI Assistant, which answers plain-language questions and cites the exact emails behind each answer.
- Thread Story, which rebuilds a conversation as a cited timeline with an optional narrative grounded only to that thread.
- Daily or weekly AI Digest: sentiment trends, urgent items and key threads for a time range.
- Anomaly detection for unusual sending patterns, off-hours timing and authentication failures, and AI Smart Filters that turn a plain-English description into a filter.
Because the index and the models are both local, the same query gives the same answer with Wi-Fi off. The on-device AI section of the home page lists these features; the point here is that none of them has a server behind it.
AIProvenance: every output says where it came from
mailin tags each AI output with the engine that produced it. A summary generated by Apple Intelligence is labeled as such; an entity list from NaturalLanguage is labeled as that; the same applies to anything produced through the optional Cloud AI mode.
This is a small feature with a practical purpose. If you are writing up findings, you can state which engine produced a given summary. If you upgrade a device and re-run an analysis, you can tell old outputs from new. And if a result looks off, you know which engine to question.
The optional Cloud AI mode, and what it sends
mailin also has a Cloud AI mode. It is opt-in and off by default, and it is disclosed clearly in the app. When you turn it on, you supply your own OpenAI or Anthropic API key; mailin does not ship with a key of its own.
Only the excerpts you choose are sent, and they go to your provider under your account. Nothing is sent to the developer, in Cloud AI mode or otherwise. Each feature that can use the cloud asks for explicit consent separately, so enabling the mode for one task does not quietly enable it everywhere.
To put that in context: the shipped app has only two things that can touch the network at all. One is App Store purchase verification through StoreKit. The other is this Cloud AI mode, if you turn it on. There are no analytics, no telemetry and no third-party SDKs, and the App Store privacy label reads "The developer does not collect any data from this app." The privacy post covers the rest of that design, and the privacy policy has the formal version.
Treat AI output as a draft, not a finding
Whichever engine produces it, AI output can be inaccurate or incomplete. Always verify important results against the underlying emails. mailin is built to make that easy: the assistant cites the messages it drew on, and Thread Story pins every narrative to the thread it came from. Use the citations.
The free tier lets you test this on up to 500 emails with three AI assistant queries and three AI smart-filter queries, and on-device sentiment and topic analysis are included. The Personal tier removes the query caps and adds unlimited AI.
FAQ
Do I need an OpenAI or Anthropic key to use mailin's AI?
No. The default engines are Apple Intelligence and Apple NaturalLanguage, which run on your device without any key. A key is needed only if you deliberately turn on the optional Cloud AI mode.
Which AI features work on macOS 14.6 or iOS 17.6?
Everything built on NaturalLanguage: sentiment, language detection and named-entity recognition. Summaries, question-answering and reasoning need Apple Intelligence, which requires macOS 26 or later or iOS 26 or later on a supported device.
Can the developer see my emails or my AI queries?
No. mailin collects nothing, and the on-device engines never send data anywhere. In Cloud AI mode, chosen excerpts go to your own provider under your own key, not to the developer.