MCP and API Guide for Digital Marketers (2026)

I manage 43 active ad accounts, 16 WordPress sites, and the analytics accounts tied to them, single-handedly. I don’t do this with a team — I do it by connecting the right tools to AI agents. This post is the list of that connection layer: which platform has an official MCP server, which ones can only read, and which can actually touch a live campaign.

Let me say this up front: this space changed completely between April and June 2026. Meta, Google and TikTok all shipped their own official servers within a few months of each other; most guides written before that are now outdated. I verified the information below as of August 1, 2026. If you’re reading this later, check the dates.

This post is part of the Agentic AI Guide series.

What MCP actually does

MCP (Model Context Protocol) is a common standard AI assistants use to connect to external systems. Roughly speaking: it turns the Google Ads API into tools an assistant can call, so when you ask “which search term wasted money last month,” the assistant pulls the report itself.

The part that matters in practice: setting up an MCP server doesn’t grant you new permissions. You get access to whatever you already have access to in the API — the server just makes that access usable by the assistant. Which means the real work isn’t installing MCP, it’s setting up the authorization behind it correctly — I’ve given that its own section below.

Ad platforms: who can do what

Google Ads. Google published its own MCP server as open source. The key limitation: read-only. It can’t pause a campaign, change a bid, or add a negative keyword. It’s built for reporting and analysis. I use it to read all 43 accounts; when a change is needed I use separate scripts that write directly to the Google Ads API.

Meta (Facebook + Instagram). The strongest one on this list. Meta put its official MCP server into open beta on April 29, 2026: 29 tools, both reading and writing. Campaign creation, budget changes, catalog operations, account diagnostics — all there. Setup is also much easier than it used to be — you sign in with your Meta Business account, no need to open a developer app or go through app review. A process that used to take days is now down to minutes.

TikTok. TikTok also announced its own server at the TikTok World event in May 2026; the Agentic Hub inside Ads Manager opened on June 30. It claims campaign planning, launching and optimization. I don’t use it actively, so I can’t say how well it really works.

LinkedIn. I couldn’t find an official MCP server. There are third-party solutions, all built on top of the LinkedIn Marketing API. If you’re doing serious B2B work, going straight to the API is still the more solid option for now.

The analytics side

Google Analytics 4. The official server is maintained by the Google Analytics team, Apache 2.0 licensed. It covers both the Data API and the Admin API, so it doesn’t just pull reports — it can also list which properties exist under which accounts. That second part is critical if you work with multiple clients: a single setup gives you reach across every property. This one is also read-only.

Google Search Console. Google doesn’t have an official MCP server. There are community packages and they work well — search analytics, index inspection, sitemap submission. I use one of them across 6 sites. But use it knowing it isn’t official: maintenance could stop any day.

A warning here: MCP directories are full of “analytics,” “ads,” “seo” packages. In an environment that has access to your server or your ad accounts, running a package with unclear maintenance is a serious risk. I don’t install packages of unclear origin — if a platform doesn’t have its own official server, I go straight to the API.

Two APIs without MCP that are still worth knowing

Google Business Profile. If you run a local business, showing up in the map pack can be worth more than ranking. It has an API, but there’s an approval gate in front of it that most people don’t know about: enabling the API in Google Cloud isn’t enough, you also have to apply to Google for access. Unapproved projects have a zero quota, so every call returns “quota exceeded” — that’s not a rate limit, it’s a closed door. I learned this by trying it with two different service accounts and user authorization.

Google Tag Manager. No ready-made MCP server, but it has an API and it works with the same credentials as GA4. Enough to read container versions and audit tags.

Three things I learned setting this up

1. The real difficulty is authorization, not setup. Setting up the server takes five minutes. The hard part is granting the service account permission at the right level. In GA4, for example, if you grant permission at the property level, you deal with every client one by one; grant it at the account level and it flows down to every property underneath. That single distinction turns a ten-client setup from half a day into five minutes.

2. “Read-only” is more useful than you’d think. Most of the time in ad management goes into figuring out what needs to change, not making the change itself. A read-access setup handles most of that job, and it does it without the risk of a wrong command breaking a live campaign.

3. A single MCP should be able to serve more than one client. Some packages read the account ID from an environment variable — meaning one setup can only watch one client. If you have five clients, you need five separate registrations. Check for this when picking a package; moving later is annoying.

How to Set Up MCP in Claude Desktop: Step by Step (No Code Required)

Up to here I’ve covered which platforms have an MCP server; now let me show concretely how it’s set up. I picked Claude Desktop as the example because it’s the simplest method that requires no terminal commands — just editing a single file.

1. Download the Claude Desktop app. Install the Mac or Windows version from Anthropic’s site and sign in with your account.
2. Go to Settings > Developer > Edit Config. This opens a configuration file called claude_desktop_config.json — plain text, you’re not writing code, just filling in a template.
3. Find the setup instructions for the MCP server you want to connect. On the GitHub page of the platform you chose — say, GA4’s official server — there’s usually a short example JSON block to add to the config file. Copy and paste it.
4. Add the service account credential. This step is exactly the authorization matter I described in “Three things I learned setting this up” — you paste a file path or API key into the config.
5. Restart Claude Desktop. A small tool icon (🔨) appears on the chat screen; clicking it lists the tools from the MCP server you connected. If you don’t see it, there’s a typo somewhere in the config file.
6. Test it. Ask a simple question like “summarize my GA4 data from last month” and see whether the assistant actually reaches the data.

These five steps follow roughly the same skeleton for nearly every official MCP server — only the platform-specific JSON block and credential in steps 3 and 4 change. Once you’ve set up one server this way, the second one goes much faster.

Where to start

If you’re a business owner managing your own ads: start with Google Analytics and Search Console. Both are read-only, both are risk-free, and both let you ask questions you’ve never looked at before.

If you want to touch your ad account: Meta’s official server is the most mature option right now. But remember it’s beta, and set it up knowing an assistant with write access can touch a live budget.

If you work with multiple clients: plan the authorization architecture before you set up anything. Which service account, at which level, with which scope — getting this right from the start is much cheaper than fixing it later.

Claude Skills vs. MCP

Everything up to here has been about MCP: how AI connects to external systems. But a second standard has become common in recent months: Claude Skills. They don’t do the same job — don’t confuse them.

MCP is a connection standard — it lets the assistant reach your Google Ads account, your GA4 property, or your WordPress database. Claude Skills is an instruction standard: it teaches the assistant in what order, with what checklist, and with what formatting rules to do a task. Without MCP, a Skill has no system to reach; without a Skill, MCP has no discipline for when and how it should be used. For example, my Google Ads MCP gives read-only access to an account, but a separate Skill file defines the order I audit an account in, which metrics I check first, and how I format the report.

So the two aren’t competitors, they’re complementary: MCP opens access, a Skill turns that access into consistent, repeatable work. I cover this in detail in What Are Claude Skills; if you want to write your own Skill, How to Write a SKILL.md is a practical starting point.

You Can Ask Me for This Work

You can ask me for this: connecting your ad and analytics accounts to AI agents with correct authorization, typically 2 hours to 2 weeks, remote, billed hourly. Scope and process: AI-Managed Google Ads.

Frequently Asked Questions

Do I need to be a developer to set up MCP?

No, but it isn’t easy either if you’re completely unfamiliar with the command line. Setup is usually just adding a few lines to a configuration file. The real difficulty isn’t the setup, it’s the Google Cloud / Meta Business authorization behind it — that’s where most people get stuck.

Can AI really manage my campaigns?

Partially. Google Ads’ official server is read-only, so nothing touches the campaign from there. Meta and TikTok do have write access, so it’s technically possible. But what I actually do isn’t “let AI manage the campaign” — I hand off the analysis and reporting, and I make the decision. A setup that lets AI decide on its own can burn through a budget fast.

Are these tools paid?

All the official MCP servers listed here are free, and most are open source. Using the Google Ads, GA4 and Search Console APIs is also free. The only thing you pay for is your AI subscription.

How do I know if my agency is using this?

Ask directly: “are you preparing reports by hand, or pulling them from the API?” If the answer isn’t clear, ask how current the report you’re getting actually is. The real benefit of these tools is speed: an audit that takes two days by hand comes out the same day with a properly set-up connection.

How long is this information valid for?

Everything in this post was verified as of August 1, 2026. This space moves fast — Meta’s server was still in beta as I wrote this, TikTok’s was two months old. If you’re reading this later, confirm with the platforms’ own documentation.

Do I need to know code to set up MCP in Claude Desktop?

No. Setup is just pasting a ready-made JSON block into a single configuration file opened from Claude Desktop’s settings — you’re not editing it, you’re copying it. The hard part isn’t code, it’s setting up the service account authorization correctly — I covered that in the step-by-step section above and in “Three things I learned setting this up.”