AI features
What the AI features do, what they send, and how to turn them on.
AI features are off until you turn them on. OpenMail uses Anthropic's API with an API key you provide. There is no OpenMail AI service in between.
Turning AI on
- Open Settings → AI Access.
- Paste your Anthropic API key and click Save key. The key goes into the macOS Keychain.
- Choose a model.
- For each account you want to use with AI, turn on Allow AI to read this account.
Automatic summaries and triage suggestions have their own per-account switches and are off by default. Turning off an account's permission deletes the AI results stored for it and cancels requests in progress.
Features
- Summaries. A streamed summary of a conversation.
- Assistant. Press ⌘J to ask questions about your mail. It searches your local mail for the accounts you allowed. Any change it suggests, such as archiving or labeling, waits for your approval.
Summaries and answers mark the message each point comes from with a number, such as [1]. Click the number, or the numbered source below the text, to open that message. Sources are listed by sender and local time.
- Compose. Rewrite or translate the text of a draft. You see the change before it goes into the draft, and the AI does not change recipients.
- Triage suggestions. Suggestions for new inbox mail, applied only after you accept them.
Automatic summaries and triage
Both run in the background for new mail, only while an OpenMail window is focused or your Mac is on power. On battery with no focused window they pause, and new mail waits until you summarize it or open the app.
- Each batch of new mail starts at most three automatic summaries and three triage suggestions.
- Automatic summaries stop after 50 a day across all accounts, and triage suggestions after 100. Each count starts again at local midnight. A request counts once Anthropic has sent text for it, because Anthropic bills it, even when it was cut off. A request that fails before any text arrives is not counted. Summaries you ask for are not counted and never capped.
- A summary that is already cached for the conversation is not requested again.
What this costs depends on the model you chose. Claude Sonnet 5.5 is the default. Claude Opus 5.5 costs about twice as much per token, and Claude Haiku 5.5 about a twentieth. An automatic summary sends up to 80 KB of message text, about 20,000 input tokens for English and more for Chinese or Japanese, and gets back at most 8,192 output tokens, thinking included. At 50 a day that is at most about 1 million input tokens and 410,000 output tokens a day. A triage suggestion reads at most 6 KB of the conversation and gets back at most 2,048 output tokens; at 100 a day that is at most about 600,000 bytes of input and 205,000 output tokens. Most requests use far fewer output tokens than these limits. Multiply by your model's prices on Anthropic's pricing page to get a daily ceiling. If that is too much, turn on automatic summaries or triage for fewer accounts or pick a smaller model.
What is sent to Anthropic
When you use an AI feature on an allowed account, OpenMail sends the question you typed and the relevant message text, subjects, and sender and recipient addresses. Requests go from your Mac to Anthropic using your API key. Anthropic's terms and privacy policy apply to that data.
OpenMail does not send:
- Message HTML, images or attachment files.
- Text extracted on your Mac from images and PDFs.
- Mail from accounts you have not allowed.
Before sending, OpenMail removes text that a message hides with CSS, such as white-on-white or zero-size text, and marks the mail as untrusted data so instructions inside a message are not followed as commands.
Anthropic bills your API key for the requests.