How I Run Meta and Google Ads With Claude and MCP

1. What MCP Actually Is

Think of MCP as a universal plug. Before it, every AI tool needed its own custom integration for every app. With MCP, a platform publishes one "MCP server" (a connector), and any AI assistant that speaks MCP can use it.

For marketers, the important part is that the big platforms now publish their own connectors. Meta has an official Ads MCP server (I covered its setup and server URL here). Google publishes official MCP servers for Google Ads and Google Analytics. Search Console is available through community connectors built on Google's official API.

Once connected, Claude does not just "know about" your ads. It can read your actual campaigns, spend, results and settings, in your accounts, with the permissions you gave it.

2. The Stack I Actually Run

Here is the setup behind everything in this post, with account details left out. I mostly work in Claude Code (Claude running on my computer), which can also read and edit my website, so an insight from the data can become a change on the site in the same session.

3. What Claude Does Inside Your Ad Accounts

These are the five jobs I actually use it for, roughly in order of how much time they save. Tap each one for a real example.

4. Three Real Examples From This Month

1. Four countries of ad costs, pulled in minutes. For my Google Ads vs Meta Ads guide, I wanted real cost numbers instead of generic benchmarks. Claude listed every Meta and Google account I can access, skipped the inactive ones, and pulled 90 days of spend, cost per click and cost per lead from the accounts with real spend in India, the UAE, Canada and Australia. The finding that went into the post: a sale in one EdTech account cost roughly 390 times a single click.

2. A measurement trap caught before it misled anyone. One Google account looked many times cheaper per lead than Meta. Before publishing that, Claude checked what each account counted as a conversion. Google was counting WhatsApp taps, calls and map directions; Meta counted only filled-in forms. Comparing them directly would have been wrong, so the post says so instead.

3. SEO fixes from Search Console data. Claude pulled 90 days of Search Console and Analytics data and found three posts ranking on page one with almost no clicks, one post indexed for months with zero impressions, and duplicate URLs splitting a page's rankings. Each became a specific fix: new titles, a rewrite and a redirect.

5. The Guardrails: How to Let AI Into Ad Accounts Safely

An AI assistant connected to live ad budgets deserves the same care as a new team member with the same access. This is the loop I follow every time. Tap each step.

Three rules I do not break

On top of the loop, these hold every time:

• Permission prompts are a feature. My setup asks before it touches a new tool. When I first asked Claude to pull client ad data as a side step of another task, the permission system stopped it until I asked for that data directly. That is exactly the behaviour you want.
• Client data stays anonymous in public. Pulling numbers is fine; publishing them is a separate decision. Anything that leaves the account (a blog post, a deck) is anonymised to niche and country unless the client has agreed otherwise.
• Credentials never go in a chat. Connectors are authorised through the platform's own sign-in screen. Nobody pastes an access token into a prompt.

6. How to Set It Up (High Level)

You do not need to be a developer, but you do need admin access to your own accounts. The broad steps:

• Pick your assistant. Claude (Desktop or Claude Code) supports MCP connectors natively.
• Connect Meta Ads. Add Meta's official Ads MCP server and sign in with the account that has access to your ad accounts. The server details are in my Meta Ads AI connectors guide.
• Connect Google Ads. Use Google's official Google Ads MCP server and authorise it with the Google account your ad accounts are shared with. Read-only access is enough for reporting and audits.
• Add Analytics and Search Console the same way if you want traffic and SEO in the same conversation.
• Start with one read-only question you already know the answer to, and check that the numbers match what you see in Ads Manager. Then build from there.
• Expect occasional re-authorisation. Some Google connections expire and need you to sign in again; it takes a minute.

7. What It Cannot Do (Yet)

It does not replace judgment about your brand, your offer or your customer. It cannot tell you whether an ad is good; only whether it performed. It relies on the tracking you set up, so if your conversions are counted wrong, its conclusions will be wrong too, confidently. And platform data lags: very recent days are often incomplete.

The skill that matters more than ever is knowing what to ask and when an answer looks wrong. That comes from running campaigns yourself, not from the AI. For the rest of the tools around this setup, see the AI tools I use at each stage of the ad workflow.