Jev Mail Classifier

Not exactly, but it will act on wrong ones. Actions are limited to tag, move, flag, unflag, mark read, mark unread and webhook, and they fire when a category's probability clears that category's own threshold, so set thresholds deliberately. run --dry-run prints what would happen without touching the mailbox.

No, they are TypeSafe's published comparison, not a measurement of this tool. The end-to-end figure comes from TypeSafe's own benchmark of an equivalent workflow: 0.114s for $0.000081 on Jev against 8.566s for $0.014 on an LLM. What this tool adds is the IMAP plumbing, the terminal UI and the action set.

That is what the cap is for. mailbox.max_emails_per_run defaults to 25, and when a run hits it the tool prints a notice and picks the rest up next time, taking the newest unprocessed mail first. Unprocessed means missing the $JevProcessed IMAP keyword rather than unread, so opening an email does not skip it.

It sends email content, and the README does not describe a zero-retention option. Each configured category becomes one independent yes/no question in a single call per email, so one message can match several categories at once. Processed mail is marked with the private IMAP keyword $JevProcessed rather than tracked in a separate database.

Clone the repo and run `./install.sh`, which sets up a virtualenv, installs the package and opens the setup wizard. Paste one Jev API key (TypeSafe, OpenRouter or Vercel AI Gateway), your IMAP login, and start adding categories; then use `./jev-mail` from the project directory.

Processed messages are marked with a private IMAP keyword, `$JevProcessed`, so there is no separate database to keep in sync. 'Unprocessed' means missing that keyword, not unread, so opening an email does not skip it, and when capped the newest unprocessed mail is classified first.

tag adds a custom IMAP keyword, move relocates the message and creates the folder if missing, flag and unflag set or clear the star, mark_read and mark_unread set or clear the Seen flag, and webhook POSTs {category, probability, subject} to a URL.