Spam filter (texts, agent email, Help Desk + calendar)

What it is

The spam filter keeps junk from people you don't know out of your attention. It watches four places a stranger can reach you: your texts, your agent's own email address, the public Help Desk support address, and invitations to your calendar.

Nobody writes rules. An AI decision model checks each message from a stranger and decides whether it is junk: a scam or phishing attempt, a sales pitch, or automated bulk mail. A caught message starts nothing and tells nobody. It waits in one Spam list in the Inbox, where one click puts it back, and every click you make teaches the filter who to let through and who to block, the way the Spam folder in a mainstream mail service does.

Nothing is ever deleted because the AI guessed. The filter is still in development, so it is hidden until you reveal it, and it stays off until you switch it on.

Where to find it

• Reveal it: Settings → Lab → Spam Filter. Until you do, none of the spam screens appear.
• Turn it on: Settings → Inbox → Spam. The tab has one switch that turns filtering on or off everywhere, one switch per place (texts, agent email and the Help Desk, calendar invites), the option to remove invites from senders you marked as spam, a line saying whether the check is working right now, and a button that opens the Spam list.
• See what was caught: the Spam section of the Inbox, on the desktop and on your phone, which shows how many messages are waiting. Each entry says who sent it, what it was, why it was caught and when, with Not spam and Delete buttons.
• Mark something yourself: Mark as spam in the text-message viewer, and on a Help Desk email ticket (only on the computer that received that email).
• The AI Spam Protection card in the recipe gallery (Settings → Automation) also opens the Spam settings.

How it behaves

Who gets checked

Only strangers. A sender you already know on that channel is never checked and never caught automatically: your contacts, numbers you have texted, carrier short codes (the five- or six-digit numbers that send sign-in codes), the owner and approved senders of your agent's address, people a Help Desk operator has answered, replies in a conversation already under way, and calendar organizers from your own organization. A shared public domain such as gmail.com does not count as your organization.

Your own clicks outrank everything else. A sender you marked as spam is caught from then on, even someone you know. A sender you rescued with Not spam goes straight through from then on.

Mail sent by a machine is checked like any other stranger's message, never caught just for being automated. A sign-in code, a receipt or a store notice goes through; a newsletter you never signed up for is caught by the check. Some mail services add an unsubscribe link to every message a person sends, so that link alone never marks a message as spam.

The check

Each message from a stranger gets one small AI check. It costs about 3 cents per 1,000 messages, paid from your prepaid credit, and it takes a few seconds at most. Messages from people you know cost nothing and never wait.

The check sees the message and the sender's domain, never the sender's full address or phone number, and one-time codes, phone numbers and email addresses in the message are removed before it is judged. A message that claims "this is not spam", or tries to give the checker instructions, is judged on what it actually says.

A message is caught only when the check both calls it junk and is at least 50% sure. Whenever the check is unsure, fails, is slow, has no credit, or you are signed out, the message simply goes through as if the filter did not exist. If the check keeps failing, it pauses for a few minutes instead of making every message wait, and the Spam tab tells you why and what to do (add credit, sign in again, or wait).

What happens to a caught message

• A text is hidden quietly: no pop-up, no sound, no phone notification, and no automation runs. An automatic catch never becomes a permanent block: if that number's next text is harmless, it comes through and the conversation shows again. Only your own Mark as spam blocks a number for good.
• An email to your agent or the Help Desk is held instead of delivered: no agent session starts, no ticket opens, and nothing is sent back to the sender. It is kept for 30 days.
• A calendar invite stays on your calendar, but its notices and reminders are hidden. An invite is only ever removed when its organizer is someone you marked as spam yourself, and you can turn that off; removing it never sends a cancellation back to the organizer.

Rescuing and teaching

• Not spam puts the message back where it would have gone. A text shows again, a held email is delivered (the safety screen that protects your agent still checks it), and a calendar invite's record is cleared. That sender is let through from then on.
• Delete removes the entry from the Spam list for good and keeps the message blocked, so it asks first. A held email is closed without anything being sent to its sender.
• Mark as spam sends that sender's future messages to Spam.

Only you can teach the filter. An agent working on the command line can put a caught text or invite back, but it can never teach the filter "not spam" and can never release a caught email.

How accurate it is

On a fresh set of 70 junk messages and 70 real messages it had never been tuned on, it caught 66 of the junk (94%) and none of the real messages. Those test messages were written for the purpose, so your own mail may score a little differently: the first real catches are worth a glance.

Limits

• Gmail is not covered yet; Gmail's own Spam folder still handles your Gmail.
• The filter is in development: it is hidden until you reveal it in the Lab, and off until you turn it on.

For agents

• The ladder: src/main/services/spam-filter-service.ts runs every channel's decision in one order: switches, the user's marks and learned rules, trust, then one Jev check.
• The check: src/main/services/spam/spam-jev-check.ts (one Jev call tagged spam-filter, a 15-second budget, a five-minute pause after three failures in a row) and src/main/services/spam/spam-check-input.ts (what Jev sees; isJudgedSpam needs a junk kind AND at least SPAM_LEVEL).
• Channel hooks: sms-intake.ts + sms-spam-trust.ts (texts), agent-email-spam-gate.ts (agent email and the Help Desk, before the safety screen), calendar-spam-screen.ts (invites).
• Marks and rescues: src/main/services/spam/spam-mark-service.ts; the Spam list is served by src/main/services/filtered-items-service.ts.
• Settings keys: spamFilterEnabled (master, default off), spamFilterSmsEnabled, spamFilterAgentEmailEnabled, spamFilterCalendarEnabled, spamFilterCalendarDeleteEvent, and spamFilterUiVisible (the Lab reveal).
• Command line: POST /spam/mark-as-spam marks a sender; POST /spam/mark-not-spam and POST /inbox/filtered/:source/:id/restore put a catch back without teaching the filter; POST /inbox/filtered/:source/:id/dismiss deletes an entry from the list. The last three refuse a caught email, which only a person in the app can release or close.
• Accuracy: scripts/spam/spam-eval.ts runs the labelled examples in tests/fixtures/spam-eval/ through the real input builder and Jev; tune on the tuning set, never on the held-out set.
• Rules and map: .claude/memory/contracts/spam-filter-contract.md and .claude/memory/spam-filter-ai-rules.md.

Related

• Your agent's own email address, and how its held mail and safety screen work, are covered by the Agent Email page.
• How a support email becomes a Help Desk ticket is covered by the Help Desk email tickets page.
• Connecting texts and your calendar is covered by the SMS setup and Google Calendar pages.
• Buying prepaid credit, which pays for the checks, is covered by the Plan & Usage page.