What is MCP? A plain-English guide for marketers who run ads

What is MCP? It stands for Model Context Protocol: an open standard, published by Anthropic, that lets an AI model such as Claude connect to outside tools and data, whether that is your ad account, a spreadsheet, or a database. For a marketer, the practical meaning is this: instead of pasting screenshots of Ads Manager into a chat, you connect the account once, and Claude reads the real numbers and, if you allow it, acts on the campaigns. This guide explains what MCP is, why it is different from copy-pasting, how the security works, and what changes for a media buyer. No code, no protocol internals.

MCP in one analogy

Think of the USB-C port. Before it, every device needed its own cable, and every cable worked with one device. USB-C is a standard: any device that supports it plugs into any port that supports it, and the two figure out what they can do together.

MCP does the same for AI models and tools. Before it, connecting Claude to your ad account meant someone writing a custom integration for that one pairing. With MCP, the ad reporting tool publishes one standard "port," Claude plugs into it, and the two negotiate what is available. The same port works for any other AI client that speaks the protocol, and the same Claude works with any other tool that publishes one.

That is why the phrase "Claude connectors" keeps showing up. A connector is what you see in Claude's settings; MCP is the standard underneath it.

The three pieces: client, server, tools

You do not need to know how the protocol works, but three words make everything else clearer.

• The client is the app you chat in. Claude, in this case.
• The server is a service that exposes something useful. An MCP server for marketing exposes your ad accounts. It runs on the tool maker's side, not on your laptop.
• The tools are the list of things the server can do, each with a name and a description. "Query Meta Ads metrics for a date range" is a tool. "Pause a campaign" is another.

When you ask Claude a question, it reads the list of tools the connected server offers, picks the one that fits, and asks the server to run it. The server calls the real API, Meta's or Google's, and sends the result back. Claude turns the result into an answer. You see the answer and, in Claude, which tool was used.

Two things follow from this. First, Claude only knows what the tools expose. If the server has no tool for a metric, Claude cannot invent it. Second, the tool maker decides which tools exist, including whether there are any that change things rather than just read them.

A screenshot versus a connected account

Most media buyers have already used Claude for ads by pasting a screenshot or an export. It works, up to a point. Here is the point.

The row that matters most in practice is the follow-up. A report never answers the first question only. CPL rose; which campaign; which ad set; since when; creative or audience. With a screenshot, each of those is a trip back to Ads Manager. With a connection, each is a line in the same chat.

How the security works, and what to ask before you connect

A connector that can read your clients' ad accounts, and maybe change budgets, deserves a few questions. The protocol itself is open and public; what it does not do is decide how a given server handles authorization. That is the server's job, so ask the server's maker.

How do I authorize it? The good answer is OAuth: you log in with the platform you already use (Facebook, Google), grant access, and the server receives a token, never your password. You can revoke that token from the platform's own settings at any time. If a server asks you to type your Facebook password into anything other than Facebook, walk away.

Which accounts does it see? You should choose. An agency with 30 accounts in Business Manager may want to connect five. The scope should be yours to set and to shrink later.

Can it change anything, and is that off by default? Reading a report and pausing a campaign are different powers. A well-built server keeps writing off until you turn it on, and keeps creating campaigns as a separate permission again.

Does it ask before acting? Every action that touches a live campaign should be confirmed in the chat before it runs. Claude shows you what it is about to do; you say yes or no.

Is there a cap and a log? A ceiling per action ("no single change may set a daily budget above US$ 500") stops a misunderstanding from becoming an expensive one. A change history, with who did it, from where, and the state before and after, is how you find out what happened when a client asks.

In Vazante, authorization is OAuth with Facebook and Google, so Vazante never sees your password. You pick which ad accounts and Pages to include. Write actions are off by default and turned on under Account and team. You set the cap per action; a request above it is refused before it reaches Meta or Google. Every action asks for confirmation in the chat. Every change is logged, including the refused ones. Creating campaigns and ads is a separate permission, also off by default.

What changes for a media buyer

Not the job. The routine.

Monday morning stops being an export. You open Claude and ask: "compare this week's spend with last week's, by campaign, for all accounts." Then: "any ad set with CPL above US$ 20?" The report in Meta Ads reporting still has the same seven pieces; you just stop assembling them by hand.

Breakdowns become questions. You do not build a placement filter in case it matters. You ask for it when a number looks odd, and drop it when the answer is boring.

Two platforms fit in one sentence. "Which brought the cheaper lead last month, Meta or Google, per account?" is one question when both accounts sit behind the same server. The comparison still needs care, and Meta and Google Ads in one dashboard explains where it goes wrong, but the mechanics are no longer the hard part.

Reading and acting sit together. If the answer is "campaign X doubled its CPL," the next line can be "pause it." With write actions on and a cap in place, Claude asks you to confirm, runs the action, and it lands in the log.

What does not change: you still need to know what to ask, which is the same skill as knowing what to put in a report. Ads Manager is still where the campaign lives; a connector complements it. And the chat is not a deliverable. The client still wants something to open, so the monthly report does not disappear; it gets built faster.

How Vazante uses MCP

Vazante is a native MCP server for marketing. The address is https://vazante.app/api/mcp. Setup in Claude: Settings → Connectors → Add custom connector → paste the address → authorize. No API key, no code. You need a Claude plan that supports connectors; the step-by-step is on the connect Meta Ads to Claude page.

The server exposes 18 tools. Eight read, eight write.

The write tools only exist for you once write actions are on. The Basic plan is read-only; Executive and above include writing. TikTok Ads reporting is available alongside Meta and Google; write actions on TikTok are not, so treat those as coming, not available.

Where to go next

If you have never connected an account, start with one. Pick the account you know best, add the connector, and ask the three questions your client asked last week. Check the answers against Ads Manager. Once they hold up, Claude for paid ads covers everything Claude can read and do from there, including how to set the cap and turn on write actions without losing sleep over it.