imbox has an assistant built in. It follows one rule: the AI writes, you decide. A draft is never sent unless you press Send, or you turn on automatic sending yourself.
AI is off until you add a key. Without one, everything else in imbox works as normal.
What leaves your server
Mail is sent to your AI provider only when you use one of these features. Nothing is sent otherwise. Each feature sends only what it needs. For example, a one-line summary sends the subject, the sender and the first 3,500 characters of the message. The complete table is on the /privacy page of your own Imbox.
None of it includes your Imbox password or the sign-in tokens of your mailboxes.
Set it up
Go to Settings → AI. Choose a provider, paste your key and pick a model.
| Provider | Notes |
|---|---|
| Anthropic | The default |
| OpenAI | |
| xAI (Grok) | |
| OpenRouter | |
| Google Gemini | |
| Any OpenAI-compatible server | Ollama, LM Studio, Groq, Mistral, a cloud gateway |
Your key is stored encrypted and is never shown again. The model name is free text, with suggestions for each provider.
Press Check features. It runs short tests against your model (a basic reply, streaming, tool calls, a full tool round trip, structured output) and tells you what works, so you know before you rely on it. Imbox also adapts to servers that are picky about their options: it switches to max_completion_tokens, drops strict and tool_choice where a server refuses them, and tries again.
Most AI features have their own switch on this page.
The assistant
The assistant is a box at the bottom of every page. Click it, or press Cmd+J, and it opens as a chat above the page. Esc or Cmd+J closes it.
- On a thread, the thread you are reading is attached as context. The suggestions become Draft a reply, Reply briefly, Reply formally, Reply warmly and Summarize.
- Anywhere, ask in plain words: what is new today, is anything waiting in the Screener, find the invoice from a sender, summarize my unread mail.
- Type
@, or press+, to attach another thread. - A draft appears as a card with Send and Open in composer.
- Earlier chats are in the menu at the top of the panel.
It can search and read threads, screen senders in or out, move mail between the Imbox and Feed, use Reply Later, Set Aside and Bubble Up, manage labels, collections, clips and bundles, look up contacts, and write drafts. It reads a thread before replying to it.
Mail is untrusted. An email is text written by a stranger, and a model can be talked into things by it. So once the assistant has read mail in a conversation, the code holds some actions back, whatever the model says:
- A draft is sent by itself only to people who are already on the thread it answers. Anything else waits for you to press Send.
- A change to many threads at once is shown on a card for you to confirm. It is not applied.
- The assistant cannot change what it remembers about you in that conversation. Say it again in a new message.
Memory
The assistant keeps a small memory that you can read, edit and clear under Settings → AI → Memory: how you write (greetings, sign-offs, length), facts about you, preferences, and notes about people. It learns from mail you send (at most twice a day, and you can turn that off), from what it does for you, and from notes you add. Drafts use it.
Other places AI appears
- A daily brief of what is waiting and what is on your calendar.
- Summaries and catch-up on a long thread, and one-line summaries under cards in the Imbox.
- Screener suggestions, and offers to move a sender's mail to the Imbox.
- Events and to-dos found in an email, as suggestions you accept or ignore.
- Finishing your sentences in a reply. Off until you switch it on under Settings → Writing. When you pause while typing a reply, grey words appear after the caret that finish the sentence. Press
Tabto add them, or keep typing and they go away. Only the end of what you have written and the start of the email you are answering are sent, and nothing is asked in the middle of a line. It is not in the phone composer yet. - Away notices. When you set an away period, AI can write the short notice. You approve the text, and what is sent later is that text.
- Answers in search. If you type a question into search, such as "did Priya approve the plan?", a short answer appears above the results with links to the messages it came from. It runs only when you press Search, and only for questions.
Next
- Agents in the app: let the assistant work a mailbox for you.
- Search and shortcuts: answers in search, and the
Cmd+Jshortcut. - Configuration: where the AI settings sit.