Connecting Your Ad Platforms to Claude Code: The Complete Guide

Part 1: What changed in 2026, and the concepts

1. Why the ad platforms shipped official MCP connections within a year

Until recently, any way to touch an ad account that wasn't the graphical interface was reserved for developers: APIs, tokens, code. The average campaign manager didn't go near it, and rightly so. That wasn't the job.

What changed in 2026 isn't the API. The API has existed for years. What changed is a new layer that sits on top of it and translates between plain language and the platform's commands. That layer is called MCP, and most of the major platforms shipped an official one in under a year. Here's when, and where to find each one's official documentation:

In short, this is no longer experimental: almost every platform you run day to day already has an official, supported connector, from the platform itself. The full breakdown for each platform, including how to connect, is in Part 3. The guide currently covers the platforms that already have an official connector, and I'll add more as they ship theirs.

Note one detail: Google built its own official MCP and released it read-only. You can't change a budget or a bid through it, only read data. That hints at where Google sees the risk, and we'll come back to it in Part 6.

2. The concepts: agents, Claude Code, MCP and OAuth

Before we connect anything, a few concepts. You don't need to understand how they work under the hood, you need to know what each one does for you, and what it means for your work.

An AI agent. A regular chat answers you. An agent acts. The difference is that an agent has hands: it pulls a report itself, cross-references data across platforms, and in some cases performs an action in the account. The implication for you is direct. The moment you connect an agent to an account, the question is no longer "what can it answer," but "what did you give it permission to do." And that's your decision, not a technical detail.

Two kinds of AI in advertising, and don't confuse them. There's AI that lives inside the platforms: Google's Smart Bidding, Meta's Advantage+‎, the algorithm that decides who sees an ad and at what price. It works for you all the time, from the inside. And there's AI that works on the platforms from the outside: an agent like Claude, which connects through MCP, reads your data, analyzes it, and in some cases changes it too. This guide is about the second kind. It doesn't replace the platform's internal AI, it gives you an external assistant that looks at everything from above, cross-references across platforms, and explains it in your own language.

Where you work: Claude Code, inside the Claude desktop app. Download the Claude app from claude.ai and install it (Mac or Windows). It's the regular app, and inside it are three working modes that appear as tabs at the top of the window (in some versions the Chat tab is labeled Home). In this guide I work mainly in Code, but it's worth knowing all three:

• Chat. The chat you already know. One question, one answer. Great for a quick check like "how much did I spend yesterday on Meta." I don't work in it in this guide, it's here for comparison.
• Cowork. You give it a goal, and it runs on its own for a few steps to a finished output: it pulls data, cross-references platforms, and returns a summary with sources. Useful for a full account audit without managing it question by question.
• Code. Despite the name, it isn't just for developers, and it's the most powerful mode. It works against a folder on your computer, remembers context over time, and runs multi-step analyses that build on each other. This is the place for day-to-day work, because you can build repeatable workflows, save prompts and templates in the folder, and come back to them. Most of the work is clicks and plain writing, no terminal.

My recommendation, and it's the path this guide takes: work in Code. Chat and Cowork stay for quick checks. You connect your ad platforms once: Meta and TikTok through Customize and then Connectors, and Google Ads and Google Analytics with a few prompts to Claude Code that handle the technical part (Part 3). A connector you've added is then available in all the modes. You don't write a single line of code yourself. A note for Windows: to enable the Code tab, install Git for Windows once (free software from gitforwindows.org, install with the defaults) and restart the app.

MCP. Short for Model Context Protocol. The name does matter, because you'll meet it on every documentation page of every platform, and it's worth knowing that something simple hides behind it: a standard that connects AI agents to external systems. Think of a standard power socket. There used to be a different plug for every device, and you needed an adapter for everything. MCP is the socket everyone finally agreed on. Thanks to it, you connect Claude to TikTok exactly the same way you connect it to Meta.

OAuth. The internet's standard sign-in mechanism, the same "Sign in with Google" you know from a thousand services. Instead of handing Claude a password, you sign in directly against the ad platform, and it grants Claude limited access whose scope you set, and that you can revoke at any time. Claude never sees your password. On most platforms this is the screen where you narrow it down to read-only. On TikTok there's no choice here, so you block it on Claude's side, as we'll see.

All the work in this guide happens in Claude Code, inside the Claude desktop app, through clicks and plain writing. Connecting Google Ads and Google Analytics happens there too: you paste a few prompts, Claude Code handles the technical side once, and you just fill in your own details. You won't write a single line of code anywhere in this guide.

And you can use other tools too. MCP is an open standard, and the same connections work with other AI assistants that support it, like ChatGPT, Copilot, or Gemini. I chose Claude here because its graphical connection is convenient for people without a technical background, and because working in Cowork and Code is especially strong for ongoing analysis. The principles are identical, and the prompts carry over as they are. What doesn't necessarily exist in other tools is the per-tool control (Tool permissions), which in this guide is the main brake. If you work with another assistant, check that it has a way to block write tools before you connect a platform whose permission screen is all-or-nothing.

Back to contents

Part 2: Setup from scratch

This part assumes you've installed nothing and never opened an AI agent in your life. If you've already worked with Claude, skip section 3. Don't skip sections 4 and 6, they're the foundation for the whole guide.

3. Install the Claude desktop app

1. Go to claude.ai, download the app for your computer (Mac or Windows), and install it like any other software.
2. Sign in with an account. Note: the Connectors area, which the whole guide relies on, is available on every plan. On a Free account you can connect one custom connector only, and on paid plans (Pro, Max, Team, Enterprise) you can connect several. If you work with client accounts, you'll want a paid account anyway, both for the number of connections, for quotas, and for terms of use that fit client data (more on that in Part 11, working with client accounts).
3. Open the app. That's it, that's the install.

One point that saves disappointment later: the connection relies on your OAuth, meaning the agent can only touch what your user can already touch. If you have read-only access to a client's account, the agent gets read-only too, and that's perfectly fine for this guide.

Tap to enlarge

4. How a connection works, and why read-only

The Meta and TikTok connections follow the same pattern, and it's worth understanding it once. Google Ads and Google Analytics connect differently, with a few prompts to Claude Code that do the work, and they get their own simple guide in their chapter. The principles we'll lay out here, read-only and per-tool permission, hold for every platform.

Inside the app, go to the Connectors area (in the Code tab through the Customize menu, and in Chat and Cowork through Settings, or just from the + button in the chat box). Click Add custom connector, paste the address the platform gives you (each platform chapter below lists the exact address), and confirm. Then a window from the ad platform itself opens, asking you to sign in to your account and approve access. It's exactly like the moment you sign in with Google to a new service and see the "do you want to give this app access" screen. It's called OAuth, and you know it from a thousand other places. Note: this approval screen belongs to the ad platform itself (Google, Meta, and so on), not to Claude, which is why it looks a little different on each platform.

This moment, approving access, is one of the most important in the guide. This screen decides what the agent will be able to do: only read, or also change. And here's my first rule, the one I don't compromise on:

On the first connection to any platform, grant read-only permission. That's my recommendation, and it's firm.

Some people connect an agent with write permission from day one. I really don't. And it's important you know it isn't required at all: you can work with an agent in read-only forever, and get enormous value out of it, without letting it touch a thing. Write permission isn't a stage you have to reach. It's an option, for anyone who's built all the brakes and wants it deliberately.

Why read-only at first? Because you don't yet know how the agent behaves on this specific account, with this structure, with this currency. A read-only period, where you ask for reports and compare them to what you see in the interface, will teach you whether you can trust it. Some platforms, like Google itself, only give read-only today anyway. They made that decision for you.

The most important control: per-tool permission, inside Claude

There's another layer of control, and you mustn't miss it. After you've connected a platform, Claude lets you set a separate permission for every tool the connection exposes. Every MCP server brings a list of tools: some only read data (get, list, search), and some change things in the account (create, update, delete, pause). On the connection's Tool permissions screen you decide what's allowed: set the read tools (usually grouped as Read-only tools) to Always allow, and set the write and create tools (Interactive tools, for example create campaign) to Blocked.

Why this is the one worth investing in: it's enforced on Claude's side, it doesn't depend on your promise to the agent or on nice phrasing in the chat, and it works even when the platform's approval screen was all-or-nothing. On TikTok, for instance, you can't choose read-only in OAuth. No problem: you connect, then block all the write tools in Claude. The agent simply doesn't see them and can't call them. An agent with no write tools available is an agent that can't burn budget, even if it makes a mistake, even if it misreads an instruction.

The simple rule: at first, anything that sounds like an action (create, update, delete, pause) is set to Blocked, and anything that only reads (get, list, search) is set to Always allow. In every platform chapter below you'll see exactly how this screen looks and which tools to block. It's worth five minutes, and it's the difference between "I nicely asked it not to touch anything" and "it's technically incapable of touching anything."

A note for teams: on Team accounts, per-tool permission control isn't always available for custom connectors. If this screen doesn't exist for you, you don't have the most important brake, and then you must not connect a platform whose approval screen is all-or-nothing (TikTok) with a user who has edit permission on the ad account. Connect it only through a user who has view-only permission on the platform, or wait until the control becomes available to you. On Meta you narrow it down on Facebook's Business Integrations screen (in the Meta chapter), and that's enough.

5. The fast track: your first 15 minutes

Before we go platform by platform, here's the shortest way to feel this working. Don't start with Google. It's the most important platform, but its connection takes a slightly longer one-time setup (a developer token, usually approved instantly, and a local install instead of the paste-address-and-approve route). It's still entirely DIY, and I wrote a simple guide for it in the Google chapter, you just don't start there. Start with Meta, whose connection is the simplest: all in the browser, no tokens and no developer.

1. The app is already installed (section 3). Go to Customize and then Connectors.
2. Add a connector with Meta's address (in the Meta chapter below), sign in and approve access in read-only permission, as I explained in section 4, and block the write tools in Tool permissions.
3. Run a small verification prompt, just to confirm the pipe is live. "Running a prompt" is simple: you paste the sentence into Claude's chat box, exactly like writing a message, and press Enter. You can do it in Hebrew or English:

These prompts use your own account details. Fill in the “My account” form once and every prompt fills itself.

And in plain language, if that's easier, you can simply write: "Which ad accounts do you see, and what were the total spend and results for each of them yesterday? Give me the account names."

If your account names came back along with numbers that look right, you're connected. Now run two or three read prompts from the Meta chapter, and cross-check every number against what you see in Ads Manager. That's it, you've started. Everything else in the guide expands on this moment.

A tip that neutralizes half the mistakes: ask the agent to show how it got to the numbers. Add the following sentence to every heavy prompt, and make it a habit. It forces the agent to state what it actually did, so you catch a wrong currency, a wrong date range, or a partial pull, before you draw a conclusion. The sentence mentions GAQL: that's Google Ads' reporting language, and we'll explain it in the Google chapter. For now just leave it in, on other platforms the agent ignores it.

6. How to write a good prompt

An agent connected to an account is fast and literal. It'll answer exactly what you asked, on whatever data it happened to pull, with the assumptions that fill the gaps you left open. The quality of the answer is set almost entirely by the quality of the prompt. A vague prompt returns a safe, reasonable, and often wrong answer. A precise prompt returns an answer you can check.

The structure of a good prompt, six parts. You don't need a paragraph for each, but each one has to come through clearly in the text:

• Context and account. Which account, which platform, which market and currency, and what business facts the agent can't see. "This is the Meta account for the Israeli market, ecommerce, the conversion that matters is a purchase." Without this it guesses the business model and picks the wrong success metric.
• The goal, the decision behind the question. Not "analyze the account" but "I want to find where money is wasted so I can decide what to pause." The goal tells the agent what's good and what's bad, and how to rank. One decision per prompt.
• The exact metrics and dimensions. Name the metrics (spend, ROAS, CPA, CTR) and the breakdown (by campaign, ad set, day, placement). A vague metric is the number-one source of a wrong answer, because platforms define metrics differently. When it matters, say which attribution window and which conversion you're counting.
• The time range. Explicit start and end dates, and the currency and region you expect back. Watch out for the conversion-lag trap: the most recent days don't have full conversions, so for CPA and ROAS either drop the last few days or ask the agent to flag that the window hasn't matured yet.
• The output format. Say exactly how you want it: a table with these columns, a ranked list, a short summary. Without this the agent tends toward prose. And ask it to show the query and the rows, so the numbers are checkable.
• Constraints. The list of "don'ts." Don't invent targets, use the targets I give you. Don't make or suggest changes, this is read-only. If a data point is missing, say so, don't estimate. State every assumption you had to make.

Do and don't, side by side:

Example: vague versus precise, on the same goal (finding wasted spend). Notice how much the answer changes between the two.

A template to copy. Fill in the brackets, delete the lines that don't apply:

Back to contents

Part 3: Platform by platform

Now we'll go platform by platform. For each one I'll write what kind of connection it is, what the agent is allowed to read, how to connect, an example prompt, and a library of example read prompts (18 to 22 per platform), starting points for your own questions. Writing, for anyone still weighing it, I've gathered in Part 6 along with the risks. A tip before you start: don't try to run 20 prompts at once. On each platform, take the first two or three as a starting point, and treat the rest as a library to come back to when you need it.

I've ordered the platforms by how easy they are to connect on your own, not by importance: Meta first, Google last. An easy win first, then the rest.

The prompts here are only examples, and the real power is yours. Before we start, a point that's easy to miss and may be the most important in the guide: every prompt you'll see ahead is just one example out of endless ones. You're not limited to the list I prepared. Once the account is connected, you can ask in plain language for almost anything you'd ask a human analyst, and get an answer in seconds.

Think for a second what that means for the business. The questions worth real money are usually the ones that take too long to answer, so you simply don't ask them. "Which audiences did I test this year, and where's the best CPA by month?" "Compare the last three months to the same period last year, by product category." "Find campaigns where frequency is rising but CTR is falling, and explain what that means." Each of these used to be a half-day pivot table, or a request to a data person. Now it's one sentence, and an answer right away.

This changes your job: less time on pulling, collecting, and pasting, more time on the questions themselves and on the decisions. The examples ahead are built to teach you to phrase a good question, not to limit you. Take them, change them, and ask what actually interests you. You're the campaign managers, and you're the only ones who know which questions are worth a fortune to you. The tool just dropped the cost of asking them to almost zero.

Before you start, a small map to save you frustration, because not all the connections are the same level of difficulty:

Fill in your account details here once, and click "Fill in my prompts." Every prompt in the guide will fill in automatically, ready to copy, and next to each prompt there's a link back to this form. Anything you don't fill in stays marked in brackets, so you know what's still left to decide based on your specific analysis. Everything is saved only on your computer, in your browser. No detail is sent to me or to anyone else, in any form.

Meta: plain language across Facebook and Instagram

In April 2026, Meta shipped an official MCP that runs on its own infrastructure, with no server of your own. You connect through the browser, with no developer app to set up and no manual tokens, which makes it one of the easiest to start with.

Official docs: Meta Ads MCP Server.

What the agent reads: reports and insights, pixel and dataset health, the account activity log.

What the agent changes: creates and edits campaigns, ad sets, and ads in plain language, manages the catalog, and sets up A/B tests and conversion lift tests.

One thing to correct if you read older guides: plenty of them claimed "there's only a third-party MCP for Meta." That's no longer true. The official MCP is real and comes from Meta itself. That said, it's still in beta, and it has rough edges. Don't run write actions on a client account before you've tested it thoroughly on a test account.

How to connect, in the browser, no code and no developer (this is the easiest connection, start here):

1. In the Claude app, go to Customize, then Connectors, and click Add (you saw this screen in the setup section).
2. In the window that opens, paste https://mcp.facebook.com/ads into the URL field. Note that the address starts with https, and that holds for every platform here.

Tap to enlarge

1. After you confirm, a Facebook ads connection screen appears. Click Connect.

1. Meta's own permissions screen opens, this is the OAuth step. Sign in to the Business account. This is where you decide what the agent can touch. There's no read-only toggle on this screen: approve it, then immediately tighten things in Facebook's settings, as explained below.

1. On the Businesses selection screen, choose Opt in to current Businesses only, and check just the relevant Business Manager, not everything. Less access, less risk. For agencies this is critical: check all of your Businesses and you pull every client into the same connection and the same chats, and the only way to walk that back later is to disconnect and reconnect. One Business per connection, and if you handle several clients, read part 11 before you connect.

Now the important part: how to force read-only. Meta gives you two layers of control, and you should use both.

First layer, on the Facebook side. Go to your business tools settings at https://www.facebook.com/settings/?tab=business_tools, find ads MCP server, and click View and edit.

In the edit view, leave only Access your Facebook ads and related stats turned on, and switch off everything else, above all Manage ads for ad accounts that you have access to. That's exactly the line between read (stats) and write (manage).

Second layer, on the Claude side. In the app, open the Facebook ads connection, and under Tool permissions block the tools that write (create ad, create campaign, and so on) and allow the read tools.

Together the two layers give you a connection that reads all the data and can't change a thing. That's exactly what you want at the start. No developer token, no developer, all of it in the browser and the app.

Plan a new campaign from history, not from scratch

Before you launch something new, it's worth asking the account what already worked. These prompts build a campaign plan from your own historical performance: what to repeat, where to start, and what already wasted money. All read-only, and the plan is a recommendation for you to approve, not an action.

20 example read prompts: Meta

TikTok: the most powerful connection, and therefore the most dangerous

In May 2026, TikTok shipped the broadest official MCP out there. It wraps roughly 400 actions from the Marketing API, more than any other server in this guide, and connecting to it doesn't ask you to manage tokens or code. Add it, sign in to your account, and go.

Official docs: How to connect to TikTok for Business MCP Server.

What the agent reads: performance reports, campaign structure, audiences, creative assets.

What the agent changes: almost everything. Creates campaigns, moves budgets, changes bids, pauses ad groups, manages creative.

And that's the problem. TikTok's official docs don't spell out fine-grained write permissions. On other platforms, like Meta, you'll find a brake built into the platform side. On TikTok there isn't one. The approval screen is broad, all-or-nothing, and you can't narrow it on TikTok's side. So you apply the brake on the Claude side, and it works well.

How to connect, in the Claude app: in Connectors, Add custom connector, paste a URL and confirm. TikTok has two URLs, and the only difference between them is how many tools load with the connection, not the permissions:

• Full (Full Disclosure), about 400 tools: https://business-api.tiktok.com/open_mcp/tt-ads-mcp-flat. This is the one TikTok's official docs recommend for Claude, and every tool loads up front.
• Progressive (Progressive Disclosure), about 40 core tools: https://business-api.tiktok.com/open_mcp/tt-ads-mcp-layer. The core tools load first, and the rest surface as needed.

I'm building a read-only connection here, so Progressive is easier to start with: fewer tools to go through when you block writes. Once you paste and confirm, TikTok's approval screen opens, and it asks for broad permissions including write, because it has no other option.

Now force read-only, on the Claude side. After connecting, open the tiktok ads connection under Tool permissions, set the read tools (Read-only tools) to Always allow, and block the write tools (Write/delete tools) as Blocked.

A note for Team accounts: if the Tool permissions screen isn't available to you, don't connect TikTok with a user who has edit access. Connect with a user who has view-only access to the TikTok account, and that's the only brake you have left.

Plan a new campaign from history, not from scratch

Before you launch something new, it's worth asking the account what already worked. These prompts build a campaign plan from your own historical performance: what to repeat, where to start, and what already wasted money. All read-only, and the plan is a recommendation for you to approve, not an action.

20 example read prompts: TikTok

Google Ads: the most important, and you connect it yourself

Google built its own official MCP, open-sourced it, and gave it just three tools, all read. Through it you can run report queries, discover field names, and see which accounts you have access to. You can't touch a budget, a bid, a keyword, or a campaign.

Official docs: google-ads-mcp on GitHub.

The report queries are written in a language called GAQL, short for Google Ads Query Language. It's Google Ads' query language, a cousin of SQL, that you use to ask the account for data: which fields, over what period, with which filters. The good news is you don't need to learn it. The agent writes the GAQL for you from a plain-language request, runs it, and returns an answer.

What the agent reads: any report you can pull from the account, at the campaign, ad group, keyword, and asset level.

What the agent changes: nothing. Through the official MCP there's no write at all.

Anyone who wants to actually make changes in Google has to drop down to the direct Google Ads API, which requires a developer, a developer token, and an approved access level. In other words, Google split this on purpose: the agent gets hands for reading, but the hand that breaks things stays behind a higher fence.

Remember that every time someone sells you "an agent that runs your campaigns on its own." The biggest ad platform in the world, the one that built the MCP with its own hands, decided not to let the agent move your money. That's not me being overly cautious, that's Google's own position.

How to connect Google Ads, with Claude Code handling the technical part. I ran this route myself, from the Google Cloud project to the first prompt that returned data, including the MCC snag you'll read about later. The idea is simple, and a campaign manager does it alone: you don't type commands. You do a few steps in the browser (create a project and generate a key), and you hand the whole technical part on your computer to Claude Code. You paste it a prompt, it shows you exactly what it's about to do, asks for approval, and runs it. Once, including publishing the app to In production in a small step we'll explain, and from then on you work in plain language. Without that publish, the connection dies every seven days. The prep here also serves Google Analytics later.

Step 1: a few steps in the browser (only you can do these)

Go to console.cloud.google.com with a Google account that has access to the ad accounts you want to read, usually the same account you use to sign in to Google Ads, the one that gets you into the manager account. (A manager account, MCC, is a top-level account in Google Ads that the ad accounts sit under. If the term is new to you, there's a full explanation in Step 2.) Google Cloud is a separate Google service for developers: you go in only to issue a key, and you won't come back to it. A point that trips a lot of people up: this Google Cloud project is just a technical "drawer" that issues an access key. It's not your ad account and not your MCC, and there's nothing to configure in it beyond an empty project.

1. Create a project. Click Select a project, then New project, give it a name, and create it. Save the Project ID that appears.

1. Enable three services (APIs). In the menu go to APIs & Services, then Library, and Enable three services: Google Ads API, Google Analytics Admin API, and Google Analytics Data API. The first is for Google Ads, and the last two are for Google Analytics, which we'll connect in the next section. Enable all three now so you don't have to come back here.

1. Create a login key (OAuth Client). In the console, go to Google Auth Platform (left menu: Overview, Branding, Audience, Clients. In older versions it still shows as APIs & Services, then OAuth consent screen, same thing). Choose External and fill in the required fields. Then go to Clients (or Credentials in the old version), Create credentials, then OAuth client ID. Under Application type choose Desktop app, give it a name, and click Create. Download the file it offers (a small text file with the key details, ending in .json) and save it somewhere you'll remember, for example C:\google-mcp\client.json (on Mac, say in your Documents folder). Copy the full path by right-clicking the file and choosing Copy as path. This file is a secret in every sense (it holds a client secret): don't email it, don't save it in a synced or shared folder, and don't paste its contents into a chat. If it leaks, delete the OAuth client in Credentials and create a new one. (If you run into a step asking you to add your email to Test users, it's optional. The Test users list only matters if you stay in Testing, and in our route you publish the app in the next step, so it isn't needed.)

1. Publish the app: Publish to In production. This is the small step that keeps the connection alive over time. Without it, it drops every seven days, and the full explanation of why is right after the connection. In the console:

• Go to Google Auth Platform and click Audience (not Overview, not Branding, and not Verification Center).
• Under Publishing status click Publish app and confirm. The status changes to In production. If it already says In production and shows a Back to testing button, leave it, you're already set.
• If Google then offers Prepare for verification or Verification Center, skip it. That's a different verification process, and you don't need it to use your own accounts.

1. Get a developer token. Here you go back from Google Cloud to Google Ads itself (the interface you know). In your manager account (MCC): Admin, then API Center. The token usually comes right away at Explorer level, which already lets you read real data. Save the token string somewhere only you can see, like a password manager, and not in a Claude chat, not in WhatsApp, not in a shared document, and not in an email to a colleague. This token, together with the MCC number, opens read access to every client account sitting under your manager account. If it leaks, revoke it on that same API Center screen and issue a new one. (If it's still at Test level, which only works on test accounts, ask that same screen for an upgrade to Explorer, which is usually granted right away.)

Step 2: what Claude Code does for you

Now the comfortable part. Claude Code does the whole technical part on your computer itself. Python, pipx, and gcloud are three free helper programs that the Google connection needs on the machine. You won't use them yourself, Claude Code installs and runs them. Open Claude Code (the Code tab in the app), and paste the following prompt. Before you paste, fill in two things: the path to the JSON file you downloaded (the one you copied with Copy as path in Step 1), and the Project ID from Google Cloud. These are two one-time setup values, not part of the day-to-day prompts, so you swap them in by hand here in the prompt, and they don't have a field in the "My account" form. If you're on Mac, change "I'm on Windows" in the prompt to "I'm on Mac." Claude Code will check what's missing on the machine, install the tools it needs, and connect you to Google. Before every action it shows exactly what it's about to run and asks for approval. Read every approval request before you click, even when it's boring. Don't set always allow on commands during setup, and if Claude Code offers to print the contents of the .env or the token in the chat, refuse and ask it to point to the file only.

When the browser opens, sign in with the same Google account from the steps above. You'll probably see a red screen titled "Google hasn't verified this app." That's completely normal and not a sign of trouble: the app you created is in Testing mode and only you use it, so Google hasn't "verified" it yet. Click Advanced, then the link to continue to your app, and approve the permissions. This explanation only holds because it's the app you created a moment ago, and its name is on the screen. If this screen pops up from somewhere else, from a link someone sent or a tool you didn't set up, don't click Advanced and don't approve. This is exactly the screen people use to steal access to Google accounts.

After you approve, you get a screen confirming the sign-in worked, and you can close it and go back to Claude Code:

Tap to enlarge

Now connect Google Ads, but first prepare the secrets. Put the token in a separate file named .env in Claude Code's working folder (why it has to be this way, later). The simplest path: ask Claude Code to create an empty file named .env with the following three lines, and fill in the values yourself in Notepad, not in the chat:

GOOGLE_PROJECT_ID=your-google-cloud-project-id
GOOGLE_ADS_DEVELOPER_TOKEN=your-developer-token
GOOGLE_ADS_LOGIN_CUSTOMER_ID=your-mcc-id-digits-only

• GOOGLE_PROJECT_ID is the Project ID from Google Cloud.
• GOOGLE_ADS_DEVELOPER_TOKEN is the token you issued in API Center.
• GOOGLE_ADS_LOGIN_CUSTOMER_ID is the critical line, and this is where most people trip up. Your manager account (MCC) number, digits only, no dashes. Without this line, any question about a specific account fails with the error "User doesn't have permission to access customer," even if everything else is set up correctly. Full explanation in the next note.

Once the .env file is ready, paste this prompt into Claude Code. Note that it holds no secret, it just points to the file:

Why a .env file specifically. The token is a password in every sense. If it sits inside a command or a config file, it can leak without you noticing: staying in your command history, getting dragged into git by accident if the folder is synced, or getting exposed if you share the folder with a colleague. A local .env file, marked in .gitignore as not-for-sharing, separates the secrets from the settings: they sit in one place, don't spill out, and it's easy to swap a token without touching anything else.

Step 3: check that it works

The simplest check: ask Claude Code a question that needs the connection. If it answers with your data, it's connected.

Did your account names come back with numbers that look right? You're connected, and you can run the 22 prompts below. Sometimes the first time it takes a few minutes for the connection to be ready, so if there's no answer right away, wait a moment and try again. If you got the error "User doesn't have permission to access customer" instead, it's almost always the missing or wrong MCC id (the GOOGLE_ADS_LOGIN_CUSTOMER_ID line in .env): fix it, and ask Claude Code to refresh the connection. And if something else is stuck, you don't have to diagnose it alone: write to Claude Code "the Google Ads connection isn't working, please check it and fix it," and it will check, diagnose, and walk you through step by step. Here too, read every action it asks you to approve, and if it offers to reconnect with broader permissions or to paste the token into the chat, say no. That's exactly the advantage of working with an agent instead of a settings screen.

Alongside Ads there's also an official Google Analytics MCP, read-only as well, for GA4 insights, and especially useful for bridging the gap between the conversions the platform reports and what GA4 sees.

Plan a new campaign from history, not from scratch

Before you launch something new, it's worth asking the account what already worked. These prompts build a campaign plan from your own historical performance: what to repeat, where to start, and what already wasted money. All read-only, and the plan is a recommendation for you to approve, not an action.

22 example read prompts: Google Ads

Google Analytics: the cross-check point across every ad platform

Alongside Ads, Google has a second official MCP, this one for Google Analytics, and it's read-only as well. This isn't another ad platform, it's the other side of the equation: what really happened on the site after the click. Every platform counts conversions on its own window and method, and the windows overlap, so if you sum up every conversion the platforms report, you end up double-counting. GA4 counts once, in one place, which is what makes it the single most important cross-check a campaign manager has.

The official documentation: google-analytics-mcp on GitHub.

What the agent reads: account and property structure, GA4 reports (metrics, dimensions, date ranges and filters, including conversions and channels), real-time reports, and funnel reports. Every tool is read-only.

What the agent changes: nothing. Read-only, enforced at the permission level (the analytics.readonly scope). It can't touch a single GA4 setting.

What it gives you: real conversions and revenue, the denominator the platforms can't agree on. The truth about channels and sources, how much traffic and revenue each source/medium actually drove. Funnels and drop-off, where users bail between landing, add-to-cart, and checkout, the part the platform never sees. And landing page quality. Put simply: the platforms tell you what they claim they did, GA4 gives you one count, de-duplicated across channels. It still under-reports (consent mode, cookie loss, cross-device gaps), and the deep business truth lives in your CRM. Two terms you'll see in every prompt here: thresholding, where GA4 hides rows with few users to protect privacy; sampling, where on large reports GA4 calculates on a sample rather than the full data. In both cases the number is incomplete, which is why you ask the agent to tell you when it happened.

How to connect Google Analytics to Claude Code. What makes this easy: you already did all the setup in the Google Ads chapter. The Google Cloud project, the OAuth key, and the gcloud sign-in already covered the Analytics permission (all three APIs were enabled together, and the analytics.readonly scope was already approved). If you haven't done that yet, run the setup in the Google Ads chapter first (steps 1 and 2 there), it's shared across both. After that, you connect GA with a single prompt to Claude Code:

That's it. There's no separate token and no approval to wait on, because GA4 runs on the same sign-in as Google Ads. Fully read-only, enforced at the scope level, and there's no way through it to change any GA4 setting.

An honest caveat: GA4 on its own won't give you true cross-platform ROAS, because it only has spend from linked Google Ads (plus whatever you imported manually), not from Meta and TikTok. It gives you revenue by channel, which is half the equation. You cross-reference the spend from the platforms themselves. The two metrics that close the loop, MER and POAS, are explained in Part 10, and there's a calculator there too.

Plan a campaign from GA4 data, before building in the ad platforms

Before you launch a campaign, let GA4 point to the audiences, channels and landing pages that already convert best. That way the plan starts from real on-site behavior, not a guess. All read-only.

18 example read prompts: Google Analytics

Back to contents

Part 4: Which platforms let Claude read, and which let it write

The table pulls together everything you need to know about each platform at a glance: whether there's an official connection, how easy it is to connect on your own, what you're allowed to read, what you're allowed to change, and what the built-in brake is, if any.

Read the table across once and notice the pattern. Google went read-only. The platforms that do open up writing do it with care that varies from one to the next. Note a subtle point: the approval that pops up before every write is Claude's, not a brake from the ad platform, and it's the same across every server. The platform that offers the broadest write access with no documented dedicated brake is TikTok, which is what makes it powerful and dangerous at once. But again, all of this is on the write side. Our focus in this guide is the other half of the table, reading, and there every platform is open. Some of the platforms in the table still have no official connection, and those, along with the workarounds, are covered in Part 7.

Back to contents