Chat that stays on your Mac: when it beats a web assistant

What does "stays on your Mac" actually mean?

Chat is a page of the notch, beside Home and Meetings: you write, a model answers, and the conversation continues from there. It is the AI Assistant's typed mode, moved out of the assistant's panel and up to the top of your screen. On its defaults it uses the built-in Gemma model, which runs on your Mac. The question, the answer and everything before them in the conversation are processed locally and sent nowhere.

It also listens to nothing. Chat never opens the microphone or the system-audio tap that the meeting modes use. If you dictate a question into it, that is ordinary on-device dictation, and the audio is never written to disk.

What does stay is the conversation itself. Chat sessions are saved in a plain file in ~/Library/Application Support/Dictera/, so you can find them again under Chat History, grouped by day and searchable. The file is not synced anywhere and not encrypted beyond what FileVault provides. If you would rather nothing was written at all, switch off Save Assistant Sessions in Settings ▸ AI Assistant, or let Keep Sessions For clear the ones you have not used in a while. The security page lists the file and its controls.

And your chats stay out of your screen share. The Chat page is drawn in a window that stays out of screen sharing and recordings by apps that respect the window privacy setting in macOS: whatever the eye in the notch is set to, a share sees at most the page's empty frame.

What does a web assistant do with the same question?

It sends the question to the provider's servers, which is how it can run a model far larger than any laptop could. The conversation is usually stored there too, attached to your account, for as long as the provider's retention policy says. Depending on the service and your settings, it may also be used to improve future models.

None of that is hidden, and for most questions none of it matters. What changes the calculation is what you paste in. "Explain this error message" is one thing. The error message with a customer's name, an internal hostname and a stack trace from code under NDA is another.

The same goes for a contract clause, a paragraph from an HR case, a line from a medical letter or a draft of something not yet announced. Each is a question that is easy to ask, and each hands someone else's information to a third party that nobody agreed to.

When is the private chat the better choice?

• The text belongs to someone else. Client documents, colleagues' messages, anything covered by an NDA or a data-handling policy. With the built-in model there is no processor to approve, because nothing is processed anywhere but your Mac.
• The question is small. Rephrase this sentence, explain this term, turn these notes into three bullet points, what does this regex match. A small model handles this kind of work well, and the answer arrives without a round trip.
• You are offline. On a plane, on a train between signals, or on a network that blocks AI services. Once the model is downloaded, Chat works with Wi-Fi off.
• You would rather not have an account. Dictera has none. There is no login and no profile, so there is no history building up under your name on anyone's server.

When is a web assistant still the better tool?

A private chat is not a smaller version of the same thing. It is a different trade, and the built-in model has limits worth knowing before you rely on it.

• Anything that needs today's information. The built-in model knows what it was trained on and cannot look anything up. News, prices, release dates and anything that changed recently belong in a tool with web search.
• Long documents. The built-in model works with a short context window, a few thousand words including its own answer. A long contract or a whole report will not fit, and when a conversation outgrows the window, Chat says so rather than quietly forgetting the start.
• Pictures. Chat can take a screenshot of a display or the front window, or accept a pasted image, up to four per question, but only a remote model with vision can read one. The on-device model has no vision, so with it selected the capture button says so at once, rather than after you have asked.
• Hard reasoning. Multi-step analysis, careful code review and long pieces of writing are where larger models earn their size. A small model answers quickly and can be confidently wrong on exactly these questions.

If the question needs any of those and contains nothing sensitive, a web assistant is the right tool, and there is no reason to pretend otherwise.

Is there anything between the two?

Yes, and it is where Chat is most useful. The engine behind Chat is chosen separately from Assist and Meeting Notes, from the built-in model, a local server or a cloud provider.

A local server keeps the privacy and raises the ceiling. Ollama, LM Studio and the other presets run a model of your choice on your own hardware, which can be far larger than the built-in one if your Mac has the memory for it. The conversation still goes nowhere else. The setup is the same one described in pointing cleanup at Ollama or LM Studio: add the server once, then pick it under Chat in the notch, in Settings ▸ AI Assistant.

A cloud provider under your own key is a middle path, not a private one. Connect one of the seven preset providers (OpenAI, OpenRouter, Groq, Mistral, DeepSeek, Together AI, xAI) and the conversation goes to that provider, under your key, on its terms. The key is kept in the macOS Keychain, and the conversation and any screenshots are sent with each request. It is still one fewer account than a web assistant, and the provider is one you chose and can check. It is just not the same promise as the built-in model, and the list of what Dictera sends treats it that way.

Because each mode has its own engine, the split can follow the work. Chat can use a large cloud model for general questions while Meeting Notes stays on the built-in model for calls you would never send anywhere. Or the other way round: meetings on a local server, Chat on the built-in model for quick, private one-offs.

How do I ask without leaving the app I'm in?

Double-tap right ⇧ from any app. The notch opens on its Chat page with the caret already in the field. Type the question or dictate it, press Return, and the answer appears under it, in the conversation already going; New Chat starts a fresh one. Esc closes the page, and a call the assistant is recording keeps running.

That shortcut is what makes the private chat a habit rather than an intention. The reason people paste sensitive text into a browser tab is rarely a considered decision. It is that the tab was already open. A field that is one double-tap away, and answers without anything leaving the Mac, is easier to reach for than a policy is to remember.

Chat also answers in the language you write in. The meeting modes are fixed to one language per session, but Chat follows whatever you type, so a question in German gets an answer in German without a setting changed.

How do I decide, question by question?

Two questions settle most cases.

1. Would I be comfortable forwarding this text to a company I have not vetted? If not, it goes to the built-in model or a local server, whatever the question is.
2. Does the answer depend on something recent, long or visual? If so, and the first answer was yes, it goes to a web assistant or a cloud model you have set up.

When both answers point the same way, the choice is easy. When they conflict, and the text is sensitive and the question is hard, the answer is usually to split the question. Ask the general part somewhere large, with the names and numbers taken out. Ask the specific part privately, with them left in.

Chat is described with the rest of the assistant on the AI Assistant page, and the engines it can use are listed on the features page.