Run client delivery from the assistant you already use
What a working morning looks like when Claude, ChatGPT or Gemini can operate your client portal over MCP, with your permissions and nothing more.
· 5 min read · Product notes
It is 8:47 on a Tuesday. You have a coffee, a laptop, and a client call at half nine that you would like to walk into knowing things. So you open the assistant you already talk to all day, the same window where you drafted yesterday's proposal, and you type: what is overdue across all clients?
The answer comes back in a few seconds. Two things. The social calendar for one retainer client slipped past its date on Friday, and a homepage revision for another is sitting in review, waiting on a named approver who has been on holiday. You ask the assistant to move the social calendar's landing date to Thursday, with the reason logged: waiting on brand photography from the client. Then you ask it to send the finished pitch deck to the third client for review. Three actions, one conversation, and you have not opened a project management tool once. You walk into the half nine call knowing exactly where everything stands.
None of this is a demo script. It is what a morning looks like when your client delivery system is something your AI assistant can operate directly, rather than a website you visit and a status you reconstruct from memory.
What is actually happening underneath
The connection is made through something called MCP, and it deserves one plain paragraph. MCP is a standard plug. It lets an assistant like Claude, ChatGPT or Gemini use a tool on your behalf, with your permissions, through a proper front door that the tool itself provides. It is not a scraper. Nothing is reading your screen, guessing at buttons, or impersonating a browser. The tool publishes a set of actions it is willing to perform, the assistant asks for one by name, and the tool decides whether your account is allowed to do it. Think of it the way you think of a card machine: a standard connector on the front, the merchant's own rules behind it.
HoopoeHQ exposes 24 of these actions over MCP. Listing work, filtering it, moving a landing date with a reason attached, creating a request from an email you paste in, sending a deliverable out for review. The assistant does not get a special tunnel into the database. It gets the same doors you use, opened with your key.
The safety question, answered properly
The moment you tell an agency owner that an AI can act inside their client workspaces, a sensible alarm goes off. Client work is confidential. Workspaces exist precisely so that one client never sees another. So it is worth being precise about how the permissions work, because the design is the reassurance.
The token is personal. You create it yourself, in your own profile, and it carries your permissions and nobody else's. When the assistant acts, it acts as you. It can see the clients you can see and touch the work you can touch, and nothing beyond that, because every action goes through exactly the same server-side permission checks as clicking the button yourself would. Workspace isolation is enforced on the server, not politely requested of the client application, so an assistant with your token is bounded the same way your browser is. If you could not do it by clicking, the assistant cannot do it by asking. There is no privileged robot account, no side channel, no special case.
That framing matters more than any feature list. The question is never whether the AI is trustworthy in the abstract. The question is whether the system treats the AI as just another client of the same rules that already protect your clients from each other. That, and mandatory 2FA for agency staff, is the whole answer. The rest is convenience.
AI drafts, people decide
There is a line worth holding, and the tooling should hold it for you: the assistant prepares, a person commits.
Inside the portal itself, AI is optional and stays in its lane. Paste a rambling client email and get a drafted ticket; one click summarises a long thread before a call. Everything works without it, and nothing it drafts becomes real until a person says so. The same principle governs the MCP side. When you asked the assistant to send that pitch deck for review this morning, what the client receives is the normal review experience: feedback pinned to the artwork, named approvers, a decision recorded with a timestamp. The review itself is a human ritual with a proper paper trail. The assistant just carried the deliverable to the door.
This is also why moving a landing date through an assistant is safe to do casually. The date is a promise the client can see, and when it moves, the client sees the new date and the reason, whoever or whatever typed the instruction. The system's honesty does not depend on your discipline in the moment. There is a longer argument about that in landing dates, not deadlines.
Setting it up takes less time than reading this post
The setup is deliberately unremarkable. You create a token in your Profile. You paste your HoopoeHQ URL into your assistant as a custom connector. Then you ask it something. That is the whole procedure; the connection runs over Streamable HTTP, which is a detail you will never need to think about again.
The more interesting change is the one that arrives over the following weeks. The Monday morning ritual of opening the board and squinting stops being a ritual, because "what needs my attention today" is now a sentence rather than a scan. The My work view still exists, with its overdue, due today and this week buckets, and on busy days you will still live in it. But an increasing amount of the routine reading, checking and shuffling happens in conversation, in the tool you already had open. If your delivery still runs on forwarded emails and memory, start with getting client work out of your inbox; the assistant is much more useful when there is a system underneath it.
Where HoopoeHQ fits
HoopoeHQ is a client delivery portal that your assistant can operate over MCP with a personal token, which means the AI can only ever do what you can do, and every client-visible promise stays honest by construction. If you already pay for Claude, ChatGPT or Gemini, the connector costs nothing extra and takes a few minutes to try. The 14-day trial needs no card.