The Atlas Field Manual

Atlas runs as an app on your iPhone. Getting going takes two connections and about five minutes: sign in with the Google account Atlas should serve, and confirm it has an AI "brain." Everything else can wait.

Signing in & becoming the owner

Atlas serves exactly one owner. The first time you open it you'll meet the sign-in screen: tick the two consent boxes, then continue with Google. On a brand-new install, the first verified Google account to sign in claims ownership — from that moment, only you can get in; any other account is turned away. Signing in with Google is your first connection: it grants Atlas access to your Calendar, Gmail, and Contacts.

The two things Atlas can't run without

A useful Atlas needs just two connections. Everything else in Chapter 2 is optional and can be added whenever you like.

• Your Google account — the sign-in itself. Lights up Calendar, Gmail, and Contacts.
• Anthropic Claude — "the brain," the reasoning model behind chat and every proactive feature. This is set once by whoever prepared your install (an ANTHROPIC_API_KEY) — there's nothing to do here yourself, but it's worth knowing Atlas is quiet without it.

The Setup Wizard

On a fresh install Atlas walks you through a short wizard the first time (you can also reopen it anytime from Settings → System → Setup Wizard). Every step has a Skip for now — nothing here is permanent, and it's all editable later.

A connection is any service you let Atlas reach into — Google, your messengers, restaurant apps, your car, your thermostat. Each one it can read becomes part of your Life Graph; each one it can act through becomes a tool. You manage them all in one place: Settings → System → Connections.

How connections work

Every connection card shows a status and a tier, and offers a "How to set this up" walkthrough with a link to where you get the credentials.

Connected · Needs attention · Not connected — and tiers Required Recommended Optional

Credentials arrive one of three ways, depending on the service: a one-click OAuth sign-in (Google, Slack, Nest), pasting an API key into the card (OpenAI, Gemini, Maps, Tessie, Vapi, Omi), or a one-time manual setup for the services that run through a Mac at home (iMessage, WhatsApp, browser bridge). You can check any connection's health with Run check, or all of them at once with Check all.

Real screens · the Connections page

Every card follows the same pattern — icon, name, tier badge, live status, what it unlocks, and a "How to set this up" walkthrough. Keys you paste are encrypted and never shown again. Swipe the row for more →

The full catalog

Fifteen services, grouped the way the app groups them. Only the first two are required; the rest light up specific abilities as you add them.

Multiple accounts & calendars

Right below the Core connections is an Accounts section for the very common case of more than one identity — a personal and a work mailbox, a shared calendar, a second Google account. Add a Google account, an Exchange (Microsoft) mailbox, a Slack workspace, a read-only calendar feed (ICS), or iCloud contacts. Each one gets a label — that's the name you'll use in chat ("send it from ROI Tech," "check my work calendar"). You can rename them, toggle scanning on or off, pick a default sender, or disconnect any of them.

Atlas is a chat screen first — you'll spend most of your time typing or talking. Everything else lives behind the dark header bar at the top, which is the same on every page.

The header, left to right

The header carries the Atlas logo (tap it to return to chat), small model indicators showing which AI is active, the microphone for talking to Atlas, and the ☰ menu. The menu button wears a red badge that sums everything waiting for you — unread notifications, pending reminders, workflows to approve, and so on.

The ☰ menu

Tapping ☰ opens a dropdown of the places Atlas keeps its work. Each row shows a red count when something's waiting there.

The Settings panel (☰ → ⚙️) is where you tune the model Atlas thinks with, how it speaks, and where you reach the deeper management pages. It's organized into six sections.

Real screens · the Settings panel

Model sub-labels ("Sonnet 4.6," "GPT 5.5," "2.5 Pro") and voices (Marin / Cedar) appear exactly as shown. Only models your install has keys for show in the switcher. Swipe for more →

Section by section

🧠 AI Model

Pick which model Atlas reasons with — Claude (Sonnet 4.6), ChatGPT (GPT 5.5), or Gemini (2.5 Pro). Only the ones your install has keys for show up. The switch actually takes effect on the chat page; elsewhere it just hints where to change it.

Voice

Voice On enables spoken replies. "Hey Atlas" turns on wake-word listening so you can start hands-free. Realtime Model (Realtime or the reasoning-heavier Realtime-2) and Realtime Voice (Marin or Cedar) apply to your next voice session.

Manage

Shortcuts to Manage People (your roster of family, friends, coworkers, vendors) and Manage Memories (everything Atlas remembers about you). Both are covered in Chapter 6.

System

The plumbing: an Observability dashboard (bridge health, tool usage, tokens, crons), Connections (Chapter 2), Calendars (mark which calendars are yours vs. the household's), and the Setup Wizard. On the chat page a Clear Chat History button also appears here.

Account

Your Owner Profile (name, pronunciation, contact, defaults), Policy (Chapter 7), and — on the iOS app only — Require Face ID each time. Your signed-in email and Sign Out live at the bottom.

Support

Feedback reports a bug or a rough edge. QA and Atlas Invite also live here but are owner/builder-only — if you received Atlas as a gift, they're hidden for you.

You don't operate Atlas with buttons and forms — you just ask, in plain language, by text or voice. Atlas figures out which tools to use. The one thing to understand first is the approval card: it's how Atlas asks permission before doing anything that leaves the app.

The approval card

Reading is always free — checking your calendar, searching a restaurant, looking up a contact. But the moment Atlas is about to act in the world — send, book, call, move, control — it stops and shows you a card with the exact details. Nothing happens until you tap Approve (or say "approve it"). Tap Reject — or "cancel it" — and it's dropped.

Twelve everyday recipes

Each recipe shows what you'd say and what Atlas does. Items marked auto run with no approval (they only read or observe); approve means you'll see a card first.

Everything Atlas does is assembled from a handful of building blocks, all sitting over the one Life Graph. Most of the time you never name them — you just ask, and Atlas picks. But four of them sound alike, and knowing which is which is the difference between "remind me" and "do it for me." This chapter settles it.

The five that get confused

Here's the whole thing in one sentence each:

The comparison, side by side

And one more thing that is none of the five: a Skill. A skill is how Atlas does something well — the steps, the pitfalls, and how it checks the result — written down once and loaded before a task it covers. It does not run on its own (that is a Playbook) and it is not a single fact (that is a Memory). Find them under Skills in Settings; anything Atlas writes for itself waits there for your approval.

The decision, in five questions

Ask them in order; the first "yes" is your answer.

Playbooks — set an automation once, then forget it

A Playbook turns a repeating pattern into a standing rule: whenever a certain kind of thing happens, do a certain thing. You author it once — and that authoring is your consent — and from then on Atlas watches your world and runs the matching plan each time the trigger fires. Atlas ships with none; nothing runs on its own until you create one.

Every playbook has two settings:

• a trigger — the kind of event that sets it off ("a booking is confirmed," "a trip is cancelled," "an external meeting is added"), optionally narrowed by keywords;
• an autonomy — Ask first (Atlas drafts the plan and waits for your approval each time) or Automatic (Atlas runs it and tells you right after — capped at three automatic runs a day so it can never run away).

You create and manage playbooks just by describing them: "Set up a playbook — whenever…". Their runs appear under Workflows, and every automatic run posts a notification, so you always know what happened.

The blocks underneath

Workflow, Watch, Reminder, and the playbooks that fire them are all kinds of Job — anything that runs over time rather than finishing in one reply. And every Job, every fact, and every action is built from four foundational blocks. You'll rarely name these, but they explain how Atlas holds together.

A Policy is a standing rule you set that governs how Atlas behaves across every request, until you change it. Where a fact says what is ("I'm gluten-free"), a Policy says what Atlas may, must, must not, and should prefer to do. It's how you hand Atlas more autonomy safely — by setting the boundaries in advance.

Core vs. your policy

Policy comes in two layers, and they're kept strictly apart:

The five kinds of rule

Every policy has exactly one effect:

Setting a policy

Two ways, and they land in the same place — the Policy page (Settings → Account → 📜):

• Just say it. "Never text Ron after midnight unless it's urgent." Atlas parses it into a structured rule and reads the parse back so you can confirm it understood.
• Author it on the page. The Policy page shows Core rules (read-only) above your editable rules, which you can add, prioritize, enable, or remove.

Real screens · the Policy page

Your rules sit below the Core rules. A rule you write can make Atlas stricter — never looser than Core.

Atlas works well out of the box, but a handful of habits make it noticeably better — fewer clarifying questions, more autonomy where you want it, and total confidence that "done" means done.

Talking to Atlas

• Be specific before a write. Atlas is built to ask when it's unsure rather than guess. Naming the who, the exact when, the where, the party size, and which account up front skips a round-trip.
• Name the account when it isn't your default. "Email Sharon" goes from your personal Google. Say "use ROI Tech" or "my work calendar" to route it elsewhere.
• Use voice for quick hands-free turns; text when there's detail to review. Voice and text share one history and the same approvals, but voice keeps replies short and won't read long lists — switch to text to read an email draft or compare slots before approving.
• Repeating a question is a signal to re-check. Atlas reads calendar, locations, and prices live at the moment you ask — it won't answer a data question from an earlier reply. Ask again and it looks again.

Actions & autonomy

• Read the approval card, then approve. It's the source of truth; what runs is exactly what it shows. "Approve it" / "cancel it" work by voice too.
• Use a workflow for chains, a single ask for one-offs. "Call X, then book, then tell me" should be one request (one approval, autonomous after). A lone send for something that needed a follow-up leaves half the job undone.
• Set standing policies to earn autonomy safely. Caps, quiet hours, and bounded grants let Atlas act more on its own because you set the limits in advance.
• Bound every grant tightly. An auto-book authorizes exactly four things — venue, date, time window, party size. A different time or party is a fresh approval. Let grants expire.

Keeping Atlas's memory clean

• Correct, don't re-state. "It's actually Y, not X" replaces the fact in place; re-stating it would create a duplicate. "Forget that" removes it.
• Facts about a person go on the person. "Sharon's assistant is Dan" belongs on Sharon's record, not general memory — that's where Atlas looks for it later.
• Fix life events and graph nodes where they're owned. Edit a detected event on the Life Events page and the graph re-syncs; graph fixes from chat point you to the real source when a node mirrors a saved person or event.

Trust & safety

• Know how Atlas avoids making things up. It only says something was done after a real result comes back, never dresses a "couldn't check" up as "unavailable," and won't describe a file that isn't actually attached. If you say something it claimed isn't there, it re-checks rather than doubling down — give it a corrected spelling if it mis-heard a name.
• Notifications are one-event-one-surface. Only genuinely time-sensitive things push to your phone; suggestions and confirmations wait in the inbox. Shape what interrupts with a policy or a mode.
• Remember what Atlas won't do — and do those yourself. It never enters your passwords or payment details into third-party sites, never moves or trades money, and never defeats security checks. Travel, events, shopping, and markets are search-and-monitor only — Atlas surfaces options and hands you the checkout; it doesn't buy.

The whole system on one page — approval rules, the vocabulary, the limits, and everything you can ask, grouped by area.

Approval at a glance

Every capability falls into one of three classes:

• auto  Reads & observes. Runs immediately, no approval — checking calendars, searching, looking up contacts and prices, setting reminders, notify-only watches.
• approve  Acts in the world. Always shows a card first — sending, booking, calling, moving events, controlling devices.
• conditional  Depends. Approval only in certain cases — e.g. an auto-book watch needs one up-front approval, then runs inside its bounds.

The vocabulary

What Atlas will not do

Everything you can ask, by area