The ad platforms' own MCP servers, and planning one budget across them

A founder spending £8,000 a month has an allocation problem more than a Meta problem or a Google problem: how much goes where, what each line is actually buying, and whether the next pound does more on search, social, a podcast or a regional TV burst. The platforms' AI connectors answer a different question, which is how to operate one account faster.

That question is worth answering, and it is also why a connector per platform does not add up to a plan. Each one sees its own inventory, its own audience vocabulary and its own reporting. Nothing in any of them compares a Meta audience with a Sky AdSmart household attribute or a cinema audience pack, and nothing stops a line too small to do anything.

Here is what each platform says its server does, on its own pages and with its own dates. Two of them can change a campaign today, three are read-only, and the rest are announced, in alpha, or not found.

None of these servers plans across platforms, and none of the platforms' own pages says it does. That is reasonable: a platform's server exists to make its own account easier to run, and moving your budget to a competitor is nobody's job there.

Three things go missing when one assistant drives several of them. There is no common audience language, so "decision makers in fintech" means something different on each platform and nothing tells you how far each translation is from what you asked for. There is no allocation logic, so the split is whatever the assistant improvises, with no floor that stops a line too small to pace. And there is no single approval record, so a change approved in one conversation is invisible to the controls on the next platform.

The channels beyond the big platforms are missing altogether. UK cinema, addressable television and most named-show podcast deals are booked through sales houses, and a person there confirms the booking whatever connector your assistant has.

A planner sits above the platforms and does the three missing jobs. AdBuyMCP compiles one sentence of audience into the targeting each of seven UK channels accepts and publishes a fidelity score for each translation, so the plan tells you how much of your audience survived the trip into each channel.

It splits the budget by goal with the product's own weights, drops any line that cannot clear its plan floor and says so, and puts every consequential step behind one approval bound to the exact action. The agent's route and the browser's route are the same route: each of the 100 MCP tools wraps one REST endpoint the browser also calls, and eleven consequential operations have no tool at all.

The platforms' servers still earn a place beside that. A read-only server is a good way to pull each platform's own reporting into your assistant next to the plan, which is where you would compare them anyway.

If you are connecting an assistant to your ad accounts this month, this order keeps the money where you can see it.

The official MCP Registry, which is in preview, holds metadata for publicly accessible MCP servers and does not support private ones. A search of it on 24 September 2026 found no first-party listing for any platform in the table above, although third-party servers for several of them were listed.

The registry lists servers that anyone can reach. AdBuyMCP's server is reachable only from an approved account, through that account's own deployment and access token, so it is not listed either.