Skip to content

Your settings

Open Settings from the account menu (top right). It holds everything about you rather than about any one project: who you are, how you sign in, your defaults, and the tokens and machines you own. It lives at /settings.

The same layout as Project settings: the menu on the left, one pane beside it, one address per section. The page heading is Settings, and each pane starts with its name, in the pane’s own head, and a sentence.

Menu Holds
Account Email (read-only), username, avatar
Password Change your own
Passkeys Sign in with your device
Defaults Your default agent, and a model and effort per agent
Notifications What tells you, by which channel, and the browsers told
Access — Agent tokens Tokens you have minted for agents
Access — API tokens Tokens for your own applications, and their webhooks
Access — Machines The computers that run agents for you
  • /settings opens on Account; the others are /settings/password, /settings/passkeys, /settings/defaults, /settings/notifications, /settings/tokens, /settings/api-tokens and /settings/runners. Older addresses redirect.
  • Pending edits keep across menus; a menu with one shows an amber dot. Today only the username goes through the save bar — 1 change not saved · Account · Discard · Save settings. A half-typed password is not a pending setting and is not kept. Everything else acts at once and says so in a snackbar.
  • Going away with an unsaved username asks first (Leave without saving?).
  • On a phone /settings is a list of menus under two headings, with a ‹ Settings link back from a pane.

Email is what you sign in with and is read-only here; an administrator changes it. Your username is derived from your email’s local part the first time you sign in (ben.moez@… → ben-moez), made unique if it collides, and it prefixes every new project’s slug (<username>-<name>) so two people can both call a project “Billing” without colliding. Change it here; nothing already created is renamed. A preview under the field shows the slug a new project would get. Pick an avatar from twelve colours and a set of glyphs.

At least twelve characters. If you lose it, an administrator issues a new one; there is no reset by email. Changing it asks for the current one.

A passkey signs you in with a fingerprint, a face or a PIN, beside the password. Adding one asks for your current password first, and needs the console on https or localhost. Your password keeps working. You can then sign in from the login page, and the terminal’s browser sign-in accepts one too.

Your default agent (Claude Code, Codex or Antigravity), and, for each agent, a model and an effort per weight (“What each weight means for …”), rather than one fallback model. These apply when nothing closer to the work says otherwise: the order is the card, then you, then the project’s default agent, then the machine. A model is only ever used with the agent it belongs to.

Which things tell you, and how. Every browser that agreed to be told is listed at the top, with the switch that turns notifications on for this one; every switch on the page acts at once and a message names what it did.

Thirteen kinds, grouped as Runs, Cards, Reviews and Teams & access, each with:

  • In app keeps a row in your own feed, the one you can come back to. It is on for every kind. Turning it off stops the rows, not the push or the mail.
  • Push sends a notification to your browsers. It carries the project and what is waiting, never the question or what an agent wrote, because a notification sits on a lock screen.
  • Email sends one mail with the same words and a link. It is off for everything until you turn it on. If the column is greyed out with Mail is not set up on this instance, ask your administrator.

You are never told about something you did yourself.

Kind Who hears it Push Email
Runs
A question is waiting on you Whoever can answer it on off
A permission request is waiting on you Whoever can answer it on off
A run you started has finished You off off
A run you started has failed You off off
Cards
A card you watch moved Those who filed, watch, are assigned to or commented on it; Merged, Shipped and Declined (with its reason) on off
Somebody commented on a card you watch The card’s followers on off
Somebody mentioned you in a comment The person mentioned on off
A card was assigned to you, or taken off you The new assignee; the previous one hears took you off on off
Feedback you filed was accepted or declined You, with the reason if one was given off off
Reviews
A branch you wrote or watch was reviewed Its author, the cards’ watchers and anyone named; also when it is marked Done or Not done on off
A review was handed to you The person it was handed to (an unassigned request is every member’s, and tells nobody) on off
Teams & access
Your team’s sprint started or completed The team’s members off off
Your access to a project or team changed You, when added to a project or a team or your role changes; for a share, the team’s members off off

Not told on purpose: being removed from a project or a team, a new issue (the Roadmap’s untriaged count says so), a card added to a sprint (its history says so), and a run you stopped yourself.

A branch reviewed notification goes to the person who offered the branch, whoever started its coding run, the card’s watchers and anyone @-mentioned in the review, but never the reviewer.

The bell in the top bar opens your unread notifications over the page you are on. The notification centre, at /notifications, is the same feed as a page, with three tabs:

Tab Shows
Unread What you have not read
All Everything kept for you
From cawdev Announcements from your administrator

Beside the tabs are a project filter and a kind filter. Rows sit under a heading for each day, with Load older at the end, and a read or unread button on every row. What you have not read carries a dot and the word New. Opening the page marks nothing read; opening a row does.

A question or a permission request that is answered, and a run notice you have looked at, mark their rows read by themselves, so the feed never says something is waiting on you when Waiting on you says it is not. A row shows only while you can still see what it is about: leave a project or a team and its rows go from your feed and your count. Read rows are kept 30 days, everything 90, and at most 1000 per person.

A token can only grant what you currently hold. Since the privilege model, that means each scope is tied to a privilege on the project — you can mint a scope only while you hold its privilege, and it is checked again at every call. A token belonging to someone who has since lost access, or whose team’s grant is paused, stops working with that access rather than quietly outliving it.

Live tokens are listed; revoked ones sit behind a disclosure. Change access… widens or narrows an existing token without minting a new one, and the secret does not change, so a runner whose config gains a project reaches it on its very next call. A new token’s secret is shown once, in a dialog with Copy token and Close.

The same rule, the same list, for a token your own application holds — a support tool that files what customers report, a dashboard that shows where a card has got to. An API token starts cawa_, holds at most three scopes (backlog:write, cards:file, cards:read) and reaches a small versioned surface of its own. Under each live token sits its webhook: the URL cawdev calls when a card moves, its signing secret shown once in a dialog (Copy secret, Close), Send a test, Rotate secret (asks first) and the last deliveries. The whole of it is the Integration API section.

Your computers that run agents — yours only, never anyone else’s, with no administrator exception. Each row shows whether it is Live or Away with a phrase such as “seen 3s ago”, what it serves and what it is driving.

Unfinished work lists any checkout a machine is holding that needs a decision:

Action Does
Release Frees the reservation only — touches no file, ends resumability
Show changes What is actually sitting uncommitted in that checkout
Stash Set it aside without losing it
Commit With a message you write
Reset… Asked twice, and says exactly what it drops

Disconnect machine (formerly Forget) retires a machine that has gone for good, behind a confirmation. Each row also shows what the machine may do without asking — its accepted tool rules and whether it accepts new ones from the console, and Allow everything, which Codex and Antigravity need before they may write. Auto-resume (per machine, off by default) picks a run back up after the usage limit lifts.

Signed-in browser sessions survive an API restart, so an upgrade does not sign you out.