Why bother
Google Ads ships an official keyword research tool, Keyword Planner, with real numbers: monthly search volume, competition level, estimated cost per click. The data is accurate because it comes straight from the source. The idea was simple: give an AI assistant access to that data over MCP (Model Context Protocol), instead of tab-switching and copying numbers by hand.
Picking an implementation
Google publishes an official MCP server for the Ads API, but in practice it only covers generic account queries — customer lists, arbitrary GAQL queries, metadata. Keyword Planner isn't mentioned anywhere in it. That meant looking at third-party open-source implementations instead. The spread is wide: from heavyweight servers with 30+ tools for campaign and budget management, down to narrow utilities with three commands. For a "just give me the keyword numbers" task, the heavyweight option would have been unnecessary risk — extra permissions for functionality that isn't needed. I settled on a compact server called google-keyword-planner-mcp — a native executable with no Python or Node.js dependency, three tools: keyword idea generation, historical metrics, and impression/click forecasting. Open source, MIT-licensed, short enough to read end to end in ten minutes.
First surprise: no developer token without a manager account
Calling the Google Ads API requires a so-called developer token. The obvious move was to grab one from an existing ad account — but the "API Center" section in Google Ads turned out to be available only from a manager account (MCC), not a regular ad account. A regular account simply doesn't show that section, with no explanation in the interface. That meant setting up a separate manager account first — free, no card required, but a step nobody warns you about in advance.
OAuth client and a one-time authorisation
From there it's the standard path for Google desktop apps: create a "Desktop app" OAuth client in Google Cloud, enable the Google Ads API. The refresh token is fetched once with a small script: it spins up a local HTTP listener, opens the authorisation page in a browser, catches the returned code, and exchanges it for tokens. After that, no repeat sign-in is needed — the server refreshes access on its own.
Second wrinkle: test-only access by default
The server started up fine, the protocol handshake worked, the tool list came back correctly — but a real query against the live ad account came back denied: a brand-new developer token only works against test accounts by default. Getting data from a real account requires a separate application for elevated access (Basic/Standard) and a review wait. Everything technical is done and working — it's this one step that stretches out the longest, and it's worth budgeting time for rather than expecting an instant result.
Bottom line
The integration itself is a couple of hours of work: download the binary, set up credentials, authorise once, register the server. Most of the time goes not into the technical steps but into Google's own admin process — a manager account that doesn't get created automatically, and a queue for elevated access. After that it's just a routine tool: the AI checks real keyword demand right in the middle of writing copy or planning an ad campaign, with no context-switching and no eyeballing it.
If you're doing something similar, budget separate time for the admin side on Google's end — not just the MCP server install itself.
