MCP is an open protocol that turns a GoFeed workspace into something an assistant can talk to. You are not buying artificial intelligence here: you connect your own, Claude or ChatGPT, and it acts inside the workspace as you, with your role and with the consent you gave when you connected. Nothing is generated on this side.
A door, and who stands on the other side of it
Almost everything sold today as AI for social media is a model hidden inside a product. You type a topic, the product calls a model it has contracted, and back comes a caption you did not write and that your subscription paid for. This is the opposite arrangement.
MCP, the Model Context Protocol, is an open standard that lets an assistant discover and use another system's tools. Whoever implements it exposes a server carrying a list of operations and a set of permission rules; whoever connects is the client. Here the server is your workspace and the client is the assistant you already pay for. The conversation does not pass through us, and neither does the model: there is not a single language model call anywhere in this layer, and that absence is the proof of the sentence above.
Hence the word door. A door does not think, does not suggest and holds no opinion about your calendar. All it does is let somebody in with an identity and a set of permissions, and record what they did while they were inside. How the connection gets made, screen by screen, is on the Claude and ChatGPT page.
The practical consequence is commercial as well as conceptual. If you switch assistants tomorrow, or if yours becomes twice as capable overnight, the workspace neither notices nor charges you for it. What you buy here is the door, not the brain.
What the assistant can reach inside a workspace
The list of operations is not a separate feature: it is the same product, exposed a different way. What your workspace already holds is what sits on the other side, and not one module more. A workspace without the CRM has no CRM tools in the list, not because they are hidden but because they do not exist for that connection.
| Area | What it can do | Who has it |
|---|---|---|
| Calendar and posts | See the week, create a piece with its time and networks, edit it | Every workspace |
| Approvals | List what is waiting, approve it or send it back with a written reason | Agency and the top creator plan |
| Inbox | Read threads, reply to messages, comments and reviews | Every workspace |
| Analytics | Per network and per piece metrics, with the aggregation rule beside each | Every workspace |
| Tasks and files | Create and move tasks, upload to Drive, find a folder | Drive everywhere, tasks in agency and the creator plan with a team |
| CRM | Companies, contacts, opportunities and a bulk CSV import | Agency only |
| Briefs and quotes | Fill in a brief, build a quote from the service catalogue | Agency only |
| Bio and Auto DM | The page links, the message rules and their delivery log | Every workspace |
Two tools are worth understanding before the rest, because everything else depends on them: search and fetch. Almost no operation accepts a name, they accept identifiers, so the assistant starts by turning "the restaurant brand" into the row that belongs to it and works from there. It is dull machinery and it is what stops a vague instruction from writing into the wrong place.
The gap in the list is worth reading too, because it says as much as the contents. The workspace subscription is not there: nothing on the other side changes a plan, adds an add-on or touches a billing detail. Neither is the administration of the workspace itself. The door is for the work, and the work is publishing, replying, measuring and keeping your clients' paperwork straight, not deciding what you pay each month. That line is drawn on purpose and it is not an oversight waiting to be corrected.
Reading never asks, writing does
The connection draws a hard line between looking and touching. Reading, which is most of what you will do, does not interrupt: you ask what is on for Thursday and it tells you. Writing stops the conversation and asks for your confirmation inside the chat before anything happens, and whatever happens is recorded under your name rather than the assistant's.
That confirmation is not an interface courtesy, it is a layer of the design. An assistant that misreads an instruction, or decides to be helpful beyond its brief, meets a question before the mistake exists. And when what gets created is a post, there is another check after it: a piece scheduled by the assistant goes through exactly the same approvals as a piece written by hand, under the rules the scheduling article sets out. The door is not a shortcut around your own process.
In an ordinary conversation the difference shows up immediately. "Show me what I have this week across the three brands" is a burst of reads that happens without you intervening and ends in a summary. "Move everything from Friday to Monday" is seven writes, and all seven ask you, with the detail of what is about to change in front of you, before any of them runs. Anyone expecting to talk at a workspace and have things quietly happen finds a slower product than they imagined, and the slowness is the point.
Permission is checked three times
Every call crosses three separate filters, and one no is enough.
- The connection's consent. When you connect you pick one workspace and you see what you are authorising. A connection is a workspace: if you work in three, you need three connections, and none of them sees anything belonging to the other two.
- Your role inside that workspace. The assistant has no account of its own and no permissions of its own. It inherits yours, and where your role does not reach, neither does it. The exact shape of that split is in roles and permissions.
- The workspace plan. A module the workspace has not contracted is unreachable even for the owner, because the check runs against the workspace's module list before it looks at who is asking.
Above all three there is an owner switch that turns the whole door off for everybody at once, and each person sees and cuts their own connections without touching anybody else's. A workspace that decides it does not want assistants does not have to persuade its team one by one.
Everything in this article happens in one place in GoFeed.
Try it freeThe three jobs that justify the connection
In practice people connect for one of three reasons, and what the three have in common is that they are volume work rather than judgement work.
The first is moving a large calendar. An agency with twenty brands spends Monday morning opening tabs; through the door you ask what looks thin this week and place the pieces that already exist where they are needed. Whoever writes the captions is still you or your assistant, and the workspace is where they land.
The second is loading data. The bulk CSV import into the CRM exists through this route and not in the interface, with progress, a warning per row and an undo. If you are arriving with a contact history from another tool, the whole route is in importing your data.
The third is heavy files. A four gigabyte video does not fit into a chat conversation, so the assistant does not try to push it through: it hands you an upload link, the file travels straight to the folder you named, and it consumes the same storage quota it would have if you had dragged it in yourself.
What this door is not
It is not an assistant. You will not find a generate button here, not now and not as a surprise in some release, and the reasoning is about the business before the technology: a product that writes for you competes with the person who hired it to write, and that is not the business we are in.
Nor is it an integration with your chat. The connection lives in the workspace rather than in a single conversation, so closing the chat does not close it and opening another one does not duplicate it. It is infrastructure, with all the dullness that implies.
And it does not replace the app. Some things are done better by looking: reading a four week grid, comparing two charts, checking how a carousel actually sits are jobs for the eye rather than for a conversation. What the door handles well is the repetitive work and the work counted in dozens, and anyone who connects it hoping never to open a browser again is due a reasonable disappointment.
And it is not a parallel way in. Everything that comes through here leaves the same trail it would have left had you done it by hand: the same row in the history, the same name, the same timestamp. The day somebody asks who scheduled that post, the answer is a person with an assistant, and the record says exactly that.
