Connecting Google Ads to Claude or ChatGPT: MCP Explained for Marketers

What MCP is, in three sentences

The Model Context Protocol describes itself as an "open-source standard for connecting AI applications to external systems". Assistants such as Claude or ChatGPT can use it to access data sources and tools, and the project site compares MCP to a USB-C port for AI applications (Source: What is the Model Context Protocol (MCP)?, accessed 30 October 2026). For you as a marketer, this means that an MCP server for Google Ads lets the assistant query your account, without you exporting reports and pasting them into the chat.

The division of roles matters. The assistant phrases questions and answers. The server decides which tools exist, which data they read and whether anything can be changed at all. So MCP is not a product but the plug; what happens behind it depends on the server.

Which Google Ads MCP servers exist

Broadly, there are three kinds. Assess every server by two separate questions: what can it read? And can it write, and if so, with what safeguard?

The server in the Google Ads GitHub account

The googleads GitHub account hosts the google-ads-mcp repository with the source code for an MCP server that works with the Google Ads API (Source: GitHub googleads/google-ads-mcp, accessed 30 October 2026). The README lists three tools:

All three are query tools; no tool for making changes is listed in the README. Setup is technical: Python with pipx, a Google Cloud project with the Google Ads API enabled, and credentials via OAuth or Application Default Credentials. The README lists a developer token as a step "if your setup requires" one; to query production accounts, it must have at least Explorer access. The README also points out: "The MCP Server will expose your data to the Agent or LLM that you connect to it." Among the clients it names are Codex and "Claude Code, Cursor, VS Code".

Servers from vendors

There are also servers from software vendors that you do not have to run yourself but connect via a URL. They differ greatly: in the number of tools, in whether they bring in data from other sources such as GA4, and above all in whether and how they are allowed to write. For every vendor, check which tools write, which limits apply and who approves a change.

Helferlain as an example

Helferlain is one of these vendor servers. The address is https://mcp.helferlain.app/mcp. 50 tools are visible to MCP clients, 49 of them read-only. One changes the ad account. For agency clients, two more are added that only maintain data in Helferlain (hiding or showing a recommendation again, writing or confirming the account profile). On top of that come the two interfaces search and fetch, through which ChatGPT Deep Research reads the initial audit and its findings; they are not counted, which is why the server's tool list has 54 entries.

The only tool that changes anything in the ad account is google_ads_mutate_with_approval. Without further input, it only performs a dry run: the change is validated by Google but not carried out. For a real change, the call needs an approval token. The app does not currently have a screen that gives you such a token for the assistant; real changes therefore go through the web app or the agency console today.

Connecting in Claude

Claude calls connected MCP servers "custom connectors". According to the Claude Help Center, they are available in Claude, Cowork and Claude Desktop, including on the Free plan; there, however, only one custom connector is possible (Source: Claude Help Center, Get started with custom connectors using remote MCP, accessed 30 October 2026).

This is how to proceed on Pro and Max:

1. Open Customize > Connectors.
2. Click "+ Add" and then "Add custom connector".
3. Enter a name and the URL of the MCP server and click "Continue".
4. Review the authentication settings and finish with "Add".

On Team and Enterprise, Owners add the connector under Organization settings > Connectors; after that, each member connects individually.

The Claude Help Center explicitly warns that custom connectors connect services that have not been verified by Anthropic. It recommends reading tool approval requests carefully, only choosing "Allow always" for servers and tools you fully trust, and disabling tools you do not need.

In Claude Code, you connect a remote server from the command line with claude mcp add --transport http <name> <url> (Source: Claude Code Docs, Connect Claude Code to tools via MCP, accessed 30 October 2026).

With Helferlain, a Helferlain window opens the first time you connect. There you choose the workspace and confirm with "Approve". The step-by-step guide is in the help article Using Helferlain in Claude, ChatGPT or Cursor.

Connecting in ChatGPT

According to OpenAI's developer documentation, you add a custom MCP server in ChatGPT under Plugins: plus button, "Add custom MCP server", enter a name and the server URL, configure authentication, confirm the risk warning with "I understand and want to continue" and select "Create as a plugin". OpenAI adds: "Account and workspace policies apply to adding and using custom MCP servers." Whether your plan or workspace allows this is therefore decided by OpenAI or your admin (Source: OpenAI Developers, Connect and test your plugin, accessed 30 October 2026). OpenAI changes interface labels fairly often; check them before you set things up.

On write actions, OpenAI says two things you should know: ChatGPT currently requires manual confirmation in any conversation before write actions can be taken. And: "It is possible for write actions to occur even if the MCP server has tagged the action as read only." (Source: OpenAI Developers, Building MCP servers for plugins and API integrations, accessed 30 October 2026). The "read only" label is therefore a statement by the server, not a guarantee from the assistant.

Connecting in Cursor, Codex and Windsurf

Development environments and command-line assistants also speak MCP. The README of the Google Ads server shows, for example, the command codex mcp add <name> --url <url> for Codex, followed by codex mcp login (Source: GitHub googleads/google-ads-mcp, accessed 30 October 2026). The exact configuration is in the documentation of each tool.

For Helferlain, you add the server to mcp.json in Cursor; the steps are in the guide for Cursor.

Five questions for your first day

In every question, name the account, the period and the goal. What a server can answer depends on its tools.

1. "Run an audit of my account and name the three most important priorities. Start with the tracking check."
2. "Show me search terms from the last 30 days with costs and no conversions."
3. "Check bid strategies and budgets. Is a budget holding back campaigns that are performing well?"
4. "Compare the last 30 days with the 30 days before and explain the biggest differences."
5. "What was changed in the account in the last two weeks?"

Double-check important figures in the account. The assistant does the wording; the figures come from the server. If the first answer shows that conversions are missing, the article Setting up and checking conversion tracking in Google Ads will help; a complete order of checks is in the Google Ads audit checklist.

What you should never allow without approval

Reading is harmless, as long as you are willing to trust the server with your data. Writing is a different matter. Approval belongs in the server, not just in the chat: clicking "Allow" in the assistant is no substitute for checking exactly what will be changed and which caps apply.

You should never allow these changes without a deliberate approval and fixed limits:

• Raising bids or switching the bid strategy, for example to Target CPA or Target ROAS; what Google requires for this is explained in the article Smart Bidding requirements
• Raising budgets or sharing budgets between campaigns
• Enabling campaigns, ad groups or ads
• Changing final URLs

Helferlain shows what this can look like in a server: anything new is only created paused, bids are capped at €10, only bid strategies with a fixed cap are accepted, final URLs must be on your own website, the daily budget has a default cap of €50, and every change applies in full or not at all.

If a change breaks a limit, the response in the assistant names a code, for example cpc_cap_exceeded, budget_cap_exceeded, mutation_shape_forbidden or r2_approval_required. Nothing is changed in that case.

Cost: tool calls per plan

With vendor servers, what often counts is not the number of questions but the number of tool calls. A single question can trigger several calls, for example first the account list, then campaign data, then a comparison.

With Helferlain, calls from the assistant count against your plan's allowance.

Prices and ad accounts per plan are listed in the article Google Ads cost 2026. What AI agents can do in Google Ads beyond this, and where approval belongs, is covered in the article AI for Google Ads 2026.

Key takeaways

• MCP is an open standard, not a product. What is possible is decided by the server.
• Reading and writing must be assessed separately. For every server, check which tools write.
• Approval belongs in the server, with fixed limits, not just in a confirmation click in the chat.
• With vendor servers, budget for tool calls, not questions.

How Helferlain connects to Claude, ChatGPT and Cursor is shown on the AI integration page.