No vaporware.

Typed, not a list: standing rules ("when I say Sharon I mean…") apply on every message, never by luck; facts about a person live in one place and show on their card; search by wording or by meaning; delete has Undo. Editable — Atlas can correct or forget a saved fact, not just append.

Written-down procedures — how to book your haircut, how you like an intro email — loaded before a task they cover and followed through their verification step. Atlas can propose one after a task it had to figure out; nothing it writes is used until you approve it.

One connected map queried in a single step; nodes are editable — rename or forget a mis-captured place or event directly from chat.

Commitments you make in your own messages are captured silently and dated. Atlas checks in casually only once one is due — and quietly marks it kept when a message or calendar event shows you followed through. Mark done, snooze, or acknowledge right from the card.

Notes and voice-meeting transcripts from a connected note-taker flow into your life events and Life Graph — so what you captured becomes context Atlas can use.

Send is approval-gated; routes to SMS (green-bubble) for non-Apple numbers via the bridge. Content search across threads. Depends on the BlueBubbles Mac bridge.

Send is approval-gated. Depends on the whatsapp-web.js Mac bridge.

Multi-account (Gmail + Outlook / Microsoft Graph). Send is approval-gated; drafts are non-sending; full-body read on demand.

Search and read messages across a connected workspace; send is approval-gated. Slack activity also feeds your life events.

Who mentioned you on work Teams, in which channel and when — relayed into your inbox as metadata by a Power Automate flow, never the message text itself. Read on request; Atlas cannot post to Teams.

The bridge reports every group chat with no name. Give one a label ("Parents (Ori)") and Atlas files what happens there — a promise made in the chat, for instance — under that label; an unlabeled group is skipped, never shown as a person.

Writes are approval-gated (Google + Microsoft Graph).

Questions about your time are answered from all your calendars, Atlas’s own record of what actually happened, and things you’ve told it — reconciled into one answer, so asking in chat, in your morning brief or in a suggestion can’t give you three. Atlas can say why it believes something happened, and says so out loud when the sources disagree rather than quietly picking one.

Cancel something in Atlas and the entry usually lingers on Google, so anyone looking at your calendar still sees a meeting that isn’t happening. Atlas notices and offers, once, to clear it — only for your own events, only where it can make the change.

Tiered by your People page. Someone listed there gets three separate cards — three days out, the day before, and the day of — each due at 9 AM Eastern and open until the occasion ends. Everyone else in your contacts gets one line in the morning brief on the day.

Say "I'll attend anyway" or "that doesn't block me" and Atlas records the decision against that occurrence. The conflict is not raised again, and every decision is reviewable on Suggestions.

Book / cancel approval-gated. Reverse-engineered API via the residential-IP browser bridge.

Book / cancel approval-gated; browser-bridge dependent.

Approval-gated (Mangomint).

Plans and times a ride — pickup, destination, and when to order so you arrive on time — then hands you a one-tap Uber or Lyft link, as a reminder at the right moment if that is later. Atlas cannot order the ride: neither app gives it API access. You tap; you pay.

Approval-gated (Vapi). Objective-evaluated with voicemail / no-answer recovery.

Answers callbacks to the number Atlas dials out from: identifies itself, takes who / why / callback number, and sends you the message. Shares no personal details and commits to nothing.

Setting the thermostat is a real device action, approval-gated.

Controlling the car is a real device action, approval-gated (Tessie).

GPS via Traccar.

Degraded: macOS Sequoia disabled Find My on the bridge Mac; Traccar is the reliable path. Treat as best-effort.

The only watch that can auto-book (Resy, via a pre-approved rule). OpenTable is notify-only.

Search + price monitoring + notify. Give it a target and it alerts on a drop or when the target is crossed; give it no target and it alerts on any move of 2% or $20 or more, up or down, at most once a day. Does NOT book — booking is out of scope.

Search + nightly-rate monitoring + notify. With a target rate it alerts on a drop or target cross; with none, on any move of 2% or $20 or more, either direction, at most once a day. Does NOT book.

Search + price / stock monitoring + notify. With a target price it alerts on a drop or target cross; with none, on any move of 2% or $20 or more, either direction, at most once a day. Does NOT buy.

Quote + threshold crossing + notify (edge-triggered). Reports prices only — does NOT trade.

Catalog search (Ticketmaster) is live. Live ticket pricing and the price watch are built but gated on a SeatGeek key in the environment; without one, Atlas says so and declines rather than guessing a price.

Expected-message monitors (e.g. a device's weekly self-test); alerts when a signal is missed.

Rendered-page fetch needs the Puppeteer Mac bridge. SSRF-guarded; treated as untrusted content.

Live read-only search across travel, events, shopping and markets. Search itself never transacts; setting a price watch on a result is the separate notify-only capability above.

Observe / judge / notify. The Executive Brain has no write tools. Nudges pass a confidence gate — low-confidence items are held back or down-ranked to a passive suggestion rather than interrupting.

Author "whenever X, do Y" rules in plain chat, then list / enable / disable / delete them. Automatic runs are capped and reported after every run.

Two layers. Atlas's own non-negotiables override every rule, tool and instruction, including yours. Under them sit your standing rules — spend caps, quiet hours, standing grants, do-not-disturb — written at /policy and enforced deterministically at the approval chokepoint, not left to the model's judgment.

Short tips about something Atlas can do that you haven't used. At most three a day, at least three hours apart, no single tip more than twice a week. Dismissable, with a history and a mute.

Who hears from Atlas lives in the owner's Google Contacts labels, never in Atlas. Joining the waitlist triggers an intro email that Atlas drafts from what the person wrote; release notes go to active users the same way. Nothing is sent until the owner approves it.

Plan → wait → act → recover, exactly-once, one approval upfront. Added steps can't expand the approved tool surface.