Google Ads MCP Server: How to Make It Truly Read-Only for Claude
If you connect Claude to your Google Ads account through Google’s official MCP server and assume it can only read, check again. In version 0.0.6 of google-ads-mcp, the default configuration exposes seven tools, and four of them write to your account.
This page shows what is switched on, how to turn it off, and how to prove the result. Everything here was tested on 19 November 2026 against google-ads-mcp 0.0.6. Later versions may differ, which is exactly why the last step is a check you can rerun.
What the default install exposes
Section titled “What the default install exposes”Starting the server with its bundled configuration and listing its tools returns this:
| Tool | Type |
|---|---|
search_search | Read: runs GAQL queries |
metadata_get_resource_metadata | Read |
customers_list_accessible_customers | Read |
labels_manage_labels | Write |
labels_manage_label_associations | Write |
assets_create_asset | Write |
assets_upload_video_asset | Write |
The server’s source notes that its mutate tools use a two-step propose-and-confirm flow, so this is not a tool that silently edits campaigns. But “the model has to confirm first” is a weaker guarantee than “the tool does not exist”, and most teams asking for read-only analysis want the second one.
Why the OAuth scope will not save you
Section titled “Why the OAuth scope will not save you”Google Analytics has a separate read-only scope. Google Ads does not. The Google Ads API publishes a single OAuth scope, https://www.googleapis.com/auth/adwords, and it covers both reading and writing. You cannot request a read-only token.
That leaves two places to enforce read-only, and you should use both.
Guardrail 1: A read-only user in Google Ads
Section titled “Guardrail 1: A read-only user in Google Ads”Create a dedicated Google account, give it Read only access in the Google Ads account, and authenticate the MCP server as that user. Whatever the server exposes, the API will reject writes from an account that is not allowed to make them.
This is the control that matters most, because it lives on Google’s side and does not depend on any configuration file staying correct.
Guardrail 2: Disable the write namespaces
Section titled “Guardrail 2: Disable the write namespaces”The server groups its tools into namespaces and reads them from a tools_config.yaml file. The bundled default enables all five. Create your own file with the write namespaces off:
namespaces: customers: true # customers_list_accessible_customers search: true # search_search metadata: true # metadata_get_resource_metadata labels: false # write tools assets: false # write toolsThen point the server at it with the GOOGLE_ADS_MCP_TOOLS_CONFIG environment variable. In a Claude Desktop configuration that looks like this:
{ "mcpServers": { "google-ads-mcp": { "command": "pipx", "args": ["run", "--spec", "google-ads-mcp==0.0.6", "google-ads-mcp"], "env": { "GOOGLE_APPLICATION_CREDENTIALS": "/absolute/path/to/credentials.json", "GOOGLE_PROJECT_ID": "your-project-id", "GOOGLE_ADS_DEVELOPER_TOKEN": "your-developer-token", "GOOGLE_ADS_MCP_TOOLS_CONFIG": "/absolute/path/to/tools_config.yaml" } } }}Two details are worth copying:
- List every namespace explicitly. When the file names its namespaces, any namespace it does not mention is treated as disabled. A future release that adds a new write namespace will not switch it on for you.
- Pin the version.
google-ads-mcp==0.0.6means an upgrade is a decision you make, not something that happens the next time Claude starts.
Guardrail 3: Verify, do not assume
Section titled “Guardrail 3: Verify, do not assume”Configuration files drift. The reliable test is to start the server the same way Claude does, ask it which tools it exposes, and fail loudly if anything unexpected appears.
import asyncio, osfrom mcp import ClientSession, StdioServerParametersfrom mcp.client.stdio import stdio_client
ALLOWED = { "search_search", "metadata_get_resource_metadata", "customers_list_accessible_customers",}
async def main(): params = StdioServerParameters( command="google-ads-mcp", args=[], env=dict(os.environ) ) async with stdio_client(params) as (read, write): async with ClientSession(read, write) as session: await session.initialize() tools = {t.name for t in (await session.list_tools()).tools} print("exposed:", sorted(tools)) return sorted(tools - ALLOWED)
extra = asyncio.run(main())raise SystemExit(f"NOT read-only: {extra}" if extra else 0)Run it with GOOGLE_ADS_MCP_TOOLS_CONFIG set and it exits cleanly with three tools. Run it against the default configuration and it stops with the four write tools named. Run it again after every upgrade.
The same check for GA4
Section titled “The same check for GA4”Google’s Analytics MCP server (analytics-mcp) is safer out of the box. Version 1.1.0 exposes ten read tools, and its one write tool only registers when the server is started with --allow-write. Keep that flag out of your configuration and give the credential only the analytics.readonly scope.
Summary
Section titled “Summary”| Layer | Google Ads | GA4 |
|---|---|---|
| Account access | Dedicated user with Read only access | Dedicated user with Viewer role |
| OAuth scope | No read-only scope exists | analytics.readonly |
| Server | Disable labels and assets namespaces | Never pass --allow-write |
| Proof | Tool-list audit after every upgrade | Tool-list audit after every upgrade |