Connect your own AI to every client

Your team can connect the AI tools you already use to every client at once, and ask across them: which clients had a bad week, how one is tracking, what changed since Friday. It's set up under More → Connections → MCP Connection in your portal.

This is a different thing from the Connected AI switch on an individual client's settings page. That one controls whether the client's own people may connect their AI to their own workspace. This one is your connection, and it isn't tied to any single client.

Connecting

Two ways, and the difference matters:

• Sign in from your AI. Any team member can do this. Open the setup page, paste its link into Claude, ChatGPT or Cursor and say "set this up", then approve the connection when it asks. The connection belongs to you and stops working the moment it's disconnected or you leave.
• Create a token. Admins only. For tools that can't sign in. A token is a long-lived secret that works from anywhere it's pasted and doesn't expire with your access, which is why it's the restricted half.

Sign in from your AI unless you have a reason not to. It's fewer steps, the credentials rotate on their own, and it's the only one of the two that puts the connection in your AI tool's own connector list. A token connection shows up in your workspace, under Connections, and nowhere else - so if you've pasted a token and then gone looking for the connection in your tool's settings, that's why it isn't there.

Neither gives you data you couldn't already see - every internal role can open every client in the portal. What changes is that your AI can read it while you work.

Naming the client

Your AI names which client each request is about. If your wording matches more than one, it will ask rather than choose - and if it can't tell at all, it will tell you so. That's deliberate: acting on the wrong client's numbers isn't a mistake you'd necessarily notice.

Ask it to list your clients first if you're not sure what it can see.

Your workspace, not just your clients

Your connection can also answer the questions the Ask button in your portal answers - see The workspace assistant:

• what is left in your workspace setup,
• which clients need attention, in the same words as your Clients page,
• which client connections have stopped or are flagged, grouped the way Connection health groups them,
• every client's report schedules, including which clients have none.

These are about your whole book, so you don't have to name a client for them, and every link points at your own portal. They are on your agency's own connection only - a client's connection never has them. The team list stays in the portal, where only admins can see it.

Opening your week

Ask what needs attention across your clients and you get one list rather than a tour of twelve dashboards. What needs a person comes first - uncontacted leads, a connection that stopped - each with its next step and a link straight to the card on that client's dashboard. Then what moved in the last week, with the reason where the client's own notes give one, and any warning the dashboard already carries about those figures. It's the read to take before deciding which account to actually open.

Two things it will tell you, and both matter:

• How many clients it covered. It sweeps up to twelve at a time. If you have more, it names the ones it checked and says the rest weren't - a list that quietly covered a third of your book would look exactly like one that covered all of it.
• Any client it couldn't read. Those are named too, because a client missing from a "needs attention" list otherwise reads as a client with nothing wrong.

You don't have to ask every time. When a conversation turns to one client, your AI is asked to check in on them first: anything new since your connection last looked at that client, or anything needing action, comes up in a line or two before the answer. If nothing is, it says nothing. And if your AI tool can run scheduled tasks, the Scheduled check across clients recipe below does the whole-book version each morning without being asked.

Setting what it checks, and how often

Under each connection on your MCP Connection page there is an Alerts and check-ins bar. Press Set up (or Edit, once it is set) and you can choose four things for that connection:

Anyone on the team can change these, the same as renaming a connection.

What a check says. It opens with what moved: each client's figures for the last 7 days against the 7 before, where the change is big enough to matter (a quarter or more), biggest first. Then what is new since the last check, with a next step where one is known, then one line for anything still open. A client whose data sheet has nothing typed in for the week is said to have nothing recorded yet, not to have fallen to zero. Ask your AI to name the client for the reading notes behind a movement before taking it at face value: a jump that is really a launch is said to be one there.

Three things are worth knowing about how they behave:

• Choosing a schedule does not start one. Only your AI tool can run a scheduled task. Saving a schedule here has your AI offer to set it up the next time you ask it what needs attention, and the page shows the exact request to paste if you would rather start it now. The request ends by running the check once while you are there: a scheduled run cannot stop to ask your permission, so allow the connector every time when your AI tool asks. The line then tells you whether it is really running, with when the last scheduled check came in, or not running yet. If you delete or change the scheduled task in your AI tool, this page cannot see that: choose Start again under the status line and it goes back to not running yet, shows the request again, and has your AI offer it the next time you ask. Changing how often does the same.
• The check says what it was narrowed to. Every across-clients answer states which clients and topics the connection is set to watch, so a short list is never mistaken for a complete one. If something needs action outside your ticks, your AI is told how many such things there are.
• Picking clients is also how you get past twelve. An across-clients check covers up to twelve clients at a time. If you manage more, a picked list makes sure the ones that matter are the ones it covers.

The same settings apply when a conversation turns to one client: the check-in for that client keeps to the topics you ticked.

Ready-made recipes

Rather than remembering the right sequence each time, your connection ships prompts - jobs you pick by name. In claude.ai they appear in the attachments or slash menu; in other tools, wherever that tool lists prompts.

"Which clients need attention" is the one to open a Monday with. It and the scheduled check across clients are the only two that aren't about a single client. It stops at a shortlist rather than a plan, on purpose: it gives you the client, the one sentence saying what's wrong or unclear, and what to look at first when you open them. Anything whose data is stale comes back in its own group - that's a connection problem, not a performance problem, and the two need different responses.

It will also refuse to add your clients' figures together or rank them by amount. Those are each platform's own numbers in each client's own currency, so a total would be meaningless. Only the return figure compares across your book.

Running a report

Ask your AI to run a report and it hands you a link - straight to that client's page with the report section already open. You press Generate; it takes a minute or two and lands as a draft for review.

It doesn't generate the report itself, and it says so rather than implying otherwise. Reports are a long job that ends with someone reading the output, so they stay where that review already happens - what the AI saves you is finding the right client and the right report.

Changing something

Nothing your AI saves is applied on its own. Every change follows the same three steps:

1. It shows you the exact wording, and which client it would land on - named in capitals, before the content, because that's the thing to check. It saves nothing at this point.
2. You confirm it - and only then does anything happen.
3. It queues for review in the portal, and stays internal to your team.

A client never sees it on their own dashboard, whatever it is. That's what makes it safe to talk about a client's account here - commercials, strategy, candour about how it's going.

When a client's material has been replaced

Tell your AI the same way - "their new brand manual replaces the old branding" - and ask it to save that. As with every save on an account-team connection, nothing happens until you agree to the wording. Once you do, the Brain checks the client's earlier records and raises a review for your team, in the Brain questions panel on the client's Manage page. It never appears on the client's own dashboard, and nothing is retired until someone on the team ticks and confirms it there. Asking a question is not enough to raise one: an account-team connection never acts on a question, and your AI will say so.

Reading in notes from a call

Paste your meeting notes and ask your AI to read them into that client. The document stays internal to your team - its title and text never appear on the client's own list - and the insights it pulls out are queued individually, each with the quote from your notes that supports it. Nothing reaches their brain until you approve it.

Go through them one at a time. Notes from an internal meeting mix things worth keeping about the client - results, deadlines, decisions - with things that must never reach them, like fees or what another client owes you. The extraction cannot tell those apart, so it proposes both and leaves the judgement to you.

Approving in bulk is the one action that would undo all of this. A real set of launch notes proposed "Agency retainer underpriced" alongside three perfectly good client facts.

Who can see what your AI did

More → Connections → MCP Connection lists every live connection with who holds it, when it was last used and how many calls it has made, and lets you disconnect any of them. Underneath, recent activity names which client each request read.

Read that list. With one connection covering every client you manage, it and disconnecting are the two controls that exist - there's no way to narrow a connection to a subset of your book.

You can rename any connection from that list too. Worth doing: signing in never asks you for a name, because the AI tool registers itself - so every connection made that way arrives called "Claude" or "Cursor". Fine with one. With a shared login, or one person connecting a laptop and a phone, the activity underneath becomes several rows saying the same word. Rename them to whoever and whatever they actually are. Renaming doesn't change what a connection can do, and a disconnected one keeps its old name so past activity still reads correctly.

Caps a client has set

If a client has capped what connected AI tools may do in their workspace, that cap applies to your connection too when it's reading that client - it's about the workspace, not about who's asking. Your AI will say so rather than failing silently.

What happened to per-client connections

Your team used to connect to one client at a time. That's been replaced: a connection now covers your whole book, so there's one to set up rather than one per client. Any per-client connection your team held has been disconnected - reconnect once under More → Connections → MCP Connection.

Clients' own connections are unaffected.