Assistants are good at the part of media buying that is reading, comparing and drafting. They are useless at it without access to the account, so most people end up copying screenshots into a chat window and pasting the answer back into Ads Manager.
An MCP server closes that loop. Koast exposes its own operations as tools at https://mcp.koast.ai/mcp, so an assistant can pull yesterday’s numbers, spot the ad set that is burning budget, draft the replacement and queue it for launch — without you moving data between two windows by hand.
It is the same engine the Koast app runs on. Every tool is generated from one declaration, which means the assistant, the command line client and Koast’s own agent are always looking at the same surface. There is no second implementation to fall behind.
Setting it up takes about a minute. The step by step instructions live in the docs: Connect Claude or ChatGPT.
What your assistant can do
Read performance
Spend, results, cost per result and delivery status at campaign, ad set or ad level. Ask for last week against the week before and get the comparison, not a CSV to squint at.
Build a launch
Campaigns, ad sets and ads assembled in a draft, with your naming conventions, tracking and targeting applied. The draft is a real Koast batch, editable in the app by anyone on your team.
Edit what is already live
Budgets, schedules, statuses and targeting on running campaigns. Changes go through the same guardrails and validation the app uses, so an assistant cannot push a configuration a human could not.
Publish, against a ceiling
Publishing is its own permission, and the publishing tools stay hidden until you set a spend ceiling on the connection. No ceiling, no publishing — however the conversation goes.
Meta and Whop, one surface
Both providers sit behind the same tools. Each brand offers exactly what its provider supports, so nothing is advertised that cannot actually run on that account.
Interactive results, not walls of text
On clients that support it, results come back as views you can read — the campaign tree, a metrics panel, a launch plan — instead of a thousand lines of JSON.
Connecting it takes about a minute
There is no key to paste and nothing to self-host. Your client registers itself and you sign in with the Koast account you already have.
Add the server
Point your client at https://mcp.koast.ai/mcp. Clients that support remote MCP servers discover everything else on their own. Full instructions.
Sign in with Koast
A normal OAuth screen in your browser. You approve the connection as yourself, so it inherits your access and nothing more.
Choose what it may do
Read, write and publish are three separate permissions. Grant read only and the connection is physically incapable of changing anything. How access works.
Pick the brand
One conversation works on one brand. Your assistant lists the brands you can reach and you choose which one is active before anything is scoped to it.
Set a spend ceiling
Only needed if you want the assistant to publish. It is a hard cap on what this one connection can commit, independent of your account budgets. Keys and limits.
Ask it something
“Which ad sets are over target CPA this week?” is a good first question. Revoke the connection any time from Settings. What you can ask for.
What it will not do
The list of tools an assistant can see is not the full list. It is filtered per connection, every time, by four things: the permission you granted, your role on the team, your plan, and what the active brand’s ad provider actually supports. A read-only connection sees 33 tools. A full read, write and publish connection on a Meta brand sees 71.
A tool that is not on the list is not a tool to route around. The server refuses the call rather than finding a neighbouring operation with the same side effect, and it says which permission would have carried it. That is the whole point of putting the boundary in the server instead of in a prompt.
Connections are visible and revocable in Koast Settings alongside your API keys, with the permissions each one holds and when it was last used. Revoking one takes effect on the next request, not at the end of a session.
All of it is documented: the MCP section of the Koast docs covers connecting an assistant, creating and revoking API keys, how access and publish limits work, what you can ask for, and what each failure means.
Questions
What is an MCP server, in plain terms?
Model Context Protocol is an open standard for connecting an AI assistant to a real system. Koast runs one at https://mcp.koast.ai/mcp. Once it is connected, your assistant stops guessing about your ad accounts and starts calling the same operations the Koast app calls.
Which assistants can connect?
Any client that speaks MCP. That includes Claude and ChatGPT today, plus editors and agent frameworks that have added support. Koast implements the standard rather than a per-client integration, so a new client works the day it ships.
Can an assistant spend my money?
Only if you let it, and only up to a number you choose. Publishing is a separate permission from reading and editing, and the two publishing tools stay hidden until a spend ceiling is set on the connection. Without that ceiling an assistant can draft an entire launch and still cannot push it live.
What can it actually see?
Whatever your own Koast login can see, never more. Every request is filtered by the permission you granted the connection, by your role, by your plan, and by what the brand's ad provider supports. A read-only connection sees 33 tools; a full read, write and publish connection on a Meta brand sees 71.
Does it work on Whop accounts too?
Yes. Koast treats Meta and Whop as two providers behind one interface, so the same tools work on either and each brand only offers what its provider actually supports.
Is there a way to use this without an assistant?
Yes. The same tools are available as a command line client, so koast get-metrics and koast list-brands work in a terminal or a CI job. One declaration serves the assistant, the CLI and Koast's own agent, so none of them can drift apart.
Connect your ad accounts to your assistant.
See the MCP server running against real accounts in a twenty minute call, or connect it yourself and ask it what happened yesterday.