Two things happened over the last year that most people writing about AI and advertising have run together into one story, and separating them is the whole point of this post.
Google shipped an official MCP server for Google Ads. Meta shipped one for its ad system roughly six months later. That is the actual news, and it has nothing to do with any particular AI product.
What MCP is, in one paragraph
The Model Context Protocol is an open standard, published by Anthropic in late 2024, describing how an AI assistant discovers a tool, authenticates against it, and calls it. Before it existed, every combination of assistant and service needed its own custom integration. With it, a service publishes one server and any compatible assistant can use it.
So when Google and Meta each published an MCP server for their ad platforms, they did not integrate with one AI company. They exposed their advertising APIs to every assistant that speaks the protocol at once.
The correction worth making
The version of this going around describes connecting Manus to Claude Code so that Manus can read your ad accounts. That is not how any of it fits together.
Manus and Claude Code are both MCP clients. They are alternatives to each other in this context, not components you connect together. Either one talks to the ads MCP server directly, as can ChatGPT, Cursor, Codex, Windsurf and Gemini. The thing doing the work is the server the platform published, and the assistant is just whichever interface you prefer.
Manus is also an odd choice to build the story around. Meta acquired it in December 2025, China's market regulator ordered the acquisition unwound in April 2026, and Meta cut Manus off from its systems that June. Whatever it is now, it is not the privileged route into Meta's ad platform.
What it actually does
Reporting is where the gain is immediate and obvious. Instead of opening Ads Manager, setting a date range, adding columns and reading a table, you ask a question in words and get the answer. Published figures put a daily check of one account at around two minutes rather than twenty, and someone reviewing five accounts each morning gets that whole routine under ten minutes.
We run paid media across Pinterest, Google, Meta and TikTok, and the honest description of that saving is that it removes the part of the job that was never the job. Nobody was adding value while waiting for a dashboard to load. The judgement about what to do with the numbers is untouched.
Write access is the part to think carefully about. These servers can create and modify campaigns, not just read them. That is a genuine capability and it is also an AI agent with permission to spend your client's money. We treat read access and write access as completely different decisions, and if you manage accounts you do not own, the second one is a conversation with the client rather than a setting you enable.
Does it replace media buyers
No, and the framing tells you something about whoever is using it.
What has been automated is retrieval. Pulling numbers, assembling comparisons, noticing that a metric moved. Those were real hours and they are largely gone, which matters most to whoever was billing for them.
What has not been automated is knowing which of fifty things that moved is worth acting on, whether a drop in return on ad spend is a creative problem or a seasonality problem or a landing page that broke, and what to test next given a budget that will not cover everything. Those decisions require knowing the business, and no protocol supplies that.
The realistic effect is that the floor rises. A buyer whose value was producing weekly reports is in trouble. A buyer whose value was judgement now has more time to apply it, and clients will increasingly notice which of the two they were paying for.
Creative is the actual bottleneck now
This is the part we have felt most directly. When analysis stops taking days, the thing standing between you and the next test is whether the creative exists.
An assistant with account access can tell you which concepts are fatiguing, which formats hold attention longest, and what the winning variants have in common, in a form you can hand to a designer as an actual brief rather than a request for more ads. That has changed the shape of our week more than the reporting speed did. The queue moved from the analyst to the designer, and that is now where we spend the planning effort.
If you want to try it
Start read-only, on one account, for a week, and compare what the assistant tells you against what you would have concluded from the dashboard yourself. You are checking whether it is reading the account the way you would, and where it is not, before you trust it with anything that changes a campaign.
Keep write access separate and deliberate. The failure mode here is not a wrong number in a report. It is a change to a live campaign that nobody reviewed.
