Agent scopes and limits

When you create a token under Settings, Agents, you choose what your own agent may do: its scopes and the projects it reaches. This page explains what each scope allows, which limits your agent works within, what always waits for you and what no agent can ever do. It also covers mandates and your setting under Settings, Agents, the two ways to let an agent make changes on its own.

Read it when you set up a token, when a mandate request arrives, or when your agent reports a refusal. To create the token itself, see Connect your own agent over MCP.

What each scope allows

The Tool reference lists every tool each scope opens.

• The "New token" dialog offers only the scopes you may hold, and every one starts checked. Uncheck the ones your agent does not need. A token with no scope is refused.
• "Projects" sets the reach: "All your projects" (every project you can open) or "Only the ones you check". The reach is fixed once the token is made. To change it, make a new token and revoke the old one.
• A token limited to named projects never reads the account, even when it holds "Read the account".
• Scopes are checked again on every request. If you lose the role a scope needs, that scope stops working on every token that holds it. The token is not revoked: if the role comes back, the scope works again. Revoke any token you no longer want.

Your agent has your limits and no more

Your agent has no plan or role of its own. It works as you.

• It reads only what you can open in the app, under your role and your project access. If you are a member with access to two projects, your agent sees those two.
• A change it prepares runs the same checks as the button you would press, at the moment you apply it. Contact research is checked against today's research allowance for the project. An outline needs a plan that includes the content plan, and a deep dive of your site needs a paid plan. See Plans and limits.
• When the account is hidden or paused after a trial or a plan ends, what is refused to you is refused to your agent.
• There is no separate allowance for agents. Its questions to Spot count toward your own hourly limit, its asks use the account's AI allowance as yours do, and its cards count toward your own open cards on the project (see Limits below).
• Reads of the whole account at once stay with owners and admins. A member's agent never sees those tools: they are missing from its list, and asking for one by name through read is answered that it is not a read it can call.
• What a member sees in the app, a member's agent reads too: how the projects you can open are doing side by side, and the plan's limits with what is used, the same numbers Settings, Billing shows every member. A token limited to named projects reads only those projects and no plan limits.
• It knows the room. start and the project brief tell your agent what is used of each limit, what the plan leaves out and what the next plan lifts, so it can say in one line when an answer is thin for that reason. It never changes your plan. See How the door answers.

What always waits for you

Without a mandate, a change your agent prepares follows your setting (below). A kind of change set to "Ask me first" becomes a change card with a "From your agent" badge, followed by the name you gave the agent when you made the token. You find these cards in two places:

• In the Spot panel on that project, drawn first under "Waiting for you".
• On Settings, Agents, under "Waiting for you", for every project.

You decide each one with "Apply" (or "Apply rule", "Apply focus"), "Edit" where the card lets you leave terms out, or "Dismiss". A card you leave expires after 24 hours, and your agent can ask again. See Change cards: apply and undo.

These changes never go alone under a mandate. The mandate card lists them under "Never alone", and each waits for you as a card whatever your setting says, except where a line below says otherwise:

• An email or a post to a publisher: it goes at once only where "Emails and posts to publishers" applies at once, and a mandate never sends one
• Triaging a reply and approving suggested keywords
• Changes to your account: creating or archiving a project, renaming the account, updating your profile, and writing to or rating the support team. Renaming the account and updating your profile follow "Tidying up" in your setting.
• A mandate itself: only you approve one
• What you confirm happened outside the platform: recording an email as delivered, a page draft as sent or a contact made elsewhere, marking a conversation won or a post as posted, recording a placement or confirming placements as live, and lifting a do-not-contact entry

Some changes wait for you each time they would widen what runs without you, even under a mandate: an edit or a rewrite of a pitch that is already live, switching on a promotion rule or automatic adds, raising the research pace, taking a line off the outreach prompt's never-say list or adding an approved claim, raising what a reply may offer a publisher, and giving a mailbox wider sending hours, a higher daily limit or a shorter gap between emails.

Client-facing words follow "Client-facing links and text": the report's commentary your agent writes, and the share links it makes, apply at once or wait for you as that row says. Links your agent makes expire in 30 days unless you say otherwise; a link that never expires always waits for you as a card.

If a mailbox needs reconnecting, your agent can only point you to Settings, Mailboxes. You press "Reconnect" yourself.

Apply at once or ask first

"Apply at once or ask first", under Settings, Agents, decides whether a change Spot or your agent prepares applies at once or waits for you as a card. It has a row for each kind of change: "Tidying up", "What you track", "Paid work", "Standing rules", "Client-facing links and text" and "Emails and posts to publishers", one column for Spot and one for your agents.

• Each row is "Apply at once", "Ask me first" or "Off". Every row starts at "Ask me first". "Off" leaves that kind of change out: Spot and your agent cannot make it, and are told so.
• A change made at once shows in the Spot panel as applied, with "Take back" for a day, and in the autopilot log. A change that cannot be taken back, such as a pipeline run that has started, says so on its card first.
• "Daily limits" sets how many changes of a kind apply at once in a day. Past the limit, the next one waits for you as a card. The day is counted in UTC.
• An owner can hold a kind of change for the whole account. An owner's hold keeps that row at "Ask me first" or "Off" for everyone, whatever each person sets.
• "Always ask first on these projects" keeps every change on a project you list as a card.
• "Set everything to ask" puts every row back at "Ask me first". It does not end a mandate.

Some changes wait for you as a card whatever a row says: a change whose card carries a warning, such as banned wording in a draft, triaging a reply, approving suggested keywords, creating or archiving a project, writing to or rating the support team, a mandate request, and what you confirm happened outside the platform. Renaming the account and updating your profile follow "Tidying up", though no mandate reaches them. Emails and posts to publishers wait for you as a card unless their row, "Emails and posts to publishers", applies at once, which takes a second press to set. Every email still passes the same send checks as yours, the do-not-contact list and one pitch per person among them, and an email your agent sends now also keeps the mailbox's sending hours and its daily limit, counted across every project that sends from it. Some always wait for you as a card, whatever a row says: resuming a pitch into a conversation or past a hold, and sending again an email the mail provider may already have delivered. A mandate never lets your agent send on its own, and an email that went cannot be taken back.

What never happens through an agent

No card, no mandate and no edit reaches these. They stay yours, in the app:

• Paying a publisher
• Billing, plans and purchases
• Publishing website pages, or deleting for good
• Members, invites, roles and connecting a mailbox
• Deleting a placement
• Changing how pages are scored, engine policy and admin

A mandate request that names any of these, or names a price or a payment, is refused before it reaches you.

Mandates

A mandate is what you let one agent do on its own, on one project, until a date. It belongs to the token the agent connected with. It names what the agent may do alone in two ways, and you can mix them:

• By name: one change, such as promoting pages, researching contacts, drafting nudges, editing a pitch draft or deciding deck cards.
• By kind: every change of one kind, the rows of the apply setting: tidying up, what you track, paid work, standing rules, or client-facing links and text. A change that is more than one kind, such as tracking new prompts (paid work and what you track), needs every kind it is.
• By the outreach preset: your agent sends mandate_propose with its preset, and the card lists the ten outreach changes by name (promoting pages, researching contacts, drafting nudges, a page's angle, recommendations, the outreach angle, editing and rewriting pitch drafts, and leaving a reply ready), each one you can uncheck, and never a send. A change it lists under "Asks you first" stays there.

Within its limits:

• Decisions a day, 1 to 100. One decision is one change, which may cover up to 200 rows.
• The paid work it may do on its own a day, counted in items: each page promoted, contact researched or draft written is one. It counts for the account like your own presses. 0 means paid work always waits for you.

Each change it makes alone applies at once, shows in the Spot panel as applied with "Take back" for a day, and is listed in the autopilot log in your terms. A change whose preview carries a warning, such as banned wording in a draft, still waits for you as a card. An edit to a live pitch always waits for you as a card, and applying it pauses the pitch first, until you resume it.

What a mandate never covers: an email or a post, a change that always asks (listed above under "Never alone"), and anything in the list of what never happens. An agent that asks for one of those alone gets it under "Asks you first" on the card.

Only a token with "Make changes in your projects" can ask for a mandate.

Approve, narrow or decline a request

1. Your agent asks. A mandate card for that project appears in the Spot panel and on Settings, Agents. If you have not asked Spot or moved your deck in the last 15 minutes, you also get one email for the request.
2. Compare what was checked (the token you made and the name you gave the agent) with what the agent only says about itself, shown as quotes. Its own words never change a term.
3. Press "Edit" to narrow it: uncheck what it may do alone, pick fewer "Decisions a day" or less paid work, or bring "Ends" forward. An edit can only narrow a request, never widen it.
4. Press "Approve mandate" or "Decline" within 24 hours. A new approval replaces the mandate already in force for that token and project.

See and end a mandate

• Each mandate in force reads as one sentence under "Mandates" on Settings, Agents and under "Agents" in the Engine room. In the Engine room, a mandate on another member's token reads "Another member's agent".
• A mandate is ended, never edited. Press "End" on your own. Owners and admins can end any mandate on the project from the Engine room. From its next request, the agent can read and ask, no more.
• While the apply setting is off, an owner holds a mandate by ending it in Settings, Agents. Setting everything to ask does not end a mandate.
• A mandate you approved before mandates could name any change reads as it did: it lets its agent decide deck cards and nothing more. To widen it, approve a new one, which ends the old one.
• What the agent decides shows in the autopilot log in your terms, for example "Your agent parked 6 pages under its mandate."

Limits

How the door answers explains what your agent is told at each of these limits.

Good to know

• To stop an agent at once, press "Revoke" on its token under Settings, Agents. It stops immediately, its mandates end with it, and this cannot be undone. A mandate request still waiting from that token can only be declined.
• To take back only its standing permission, end the mandate and keep the token.
• Reads are not logged one by one. "Last used" on each token shows when an agent last used it.