Google Ads MCP server: Google's official server vs a hosted one

The Google Ads API has existed for years. MCP, the Model Context Protocol, is what lets Claude, ChatGPT, Claude Code or Cursor call it from a sentence. A Google Ads MCP server exposes tools; the client lists them, the assistant picks one from your question, the server calls the API and the answer comes back in the conversation.

Two kinds exist. Google's own server is open source and runs where you put it. A hosted server, such as Adrails, runs at a fixed address and you sign in to it. The tools, the setup and what happens to a change differ.

Google's server was announced on the Google Ads Developer Blog in October 2025 and is maintained by the Google Ads API team under the Apache-2.0 license. It exposes three tools. list_accessible_customers returns the account IDs your credentials reach. search runs a query in GAQL, the Google Ads Query Language, so the assistant writes the query from your question and gets rows back. get_resource_metadata tells it which fields a resource has, so it can write a correct query.

It is read-only in its current release, as Google's documentation states. It cannot change a bid, pause a campaign or create an ad. What it reads is wide: any resource and any field the API exposes, including ones no packaged tool covers, such as auction insights or change history.

A Google Cloud project, an OAuth client or a service account, and credentials on the machine that runs the server. Locally it runs over stdio from pipx, for a client such as Claude Code, Cursor or VS Code. For a team, you build the image, deploy it on Cloud Run and add Firestore or Redis so OAuth tokens survive a restart. Developer tokens, once the gate to the API, are no longer required since 9 September 2026 according to Google's documentation.

Each person who uses it locally repeats the setup, or the team shares the Cloud Run address. Someone owns the deployment, its credentials and its updates.

A hosted server moves the setup to a sign-in. You add one address to the client, it registers itself, and a browser page asks which workspace and which permissions. Nobody creates a Cloud project or keeps a service running. The same address works in Claude, ChatGPT, Claude Code and Codex, for everyone on the team.

It also changes what a tool can do. Because the server holds the connection to Google Ads, it can expose writes, and decide how a write happens: straight away, or as a proposal that someone applies.

performance_period reads one account over an exact date range at the account, campaign, ad group or ad level, with an optional comparison period. campaigns_structure returns how the account is built as of the last sync. google_keyword_planner returns monthly searches and top-of-page bids for a keyword list, or new ideas from a few words or a page, for the countries and languages you name. google_ads_library and targeting_options look up the account's images, audiences and the real IDs of locations and languages.

google_ads_build prepares a new Search, Display, Demand Gen or Performance Max campaign with its ad groups, keywords and ads, or its asset groups. google_ads_edit changes ad texts, keywords, names, dates, target CPA or ROAS and asset group texts. campaigns_set_status pauses or resumes, and campaigns_change_budget moves a campaign budget by at most 30% in one change. Each comes back with a link to apply it in Adrails, which reads Google Ads again first. A proposal expires after 30 minutes, and an applied change can be undone.

Google Ads opens once Google finishes verifying Adrails. Connect Meta for now. A workspace that already connected Google Ads can use these tools today.

This is the real trade. GAQL answers any question on any field, including the ones nobody packaged, and the assistant writes the query for you. It also means a wrong field name or a wrong date filter returns a wrong figure that looks right. Plain inputs answer the recurring questions in a fixed shape, with the figures the agents use, and put store revenue from Shopify or Stripe next to spend when the workspace has it. They do not answer a question nobody wrote a tool for.

A team with a developer at hand and only reads to do can run Google's server. A team that wants Keyword Planner, writes with a review step, Meta in the same connection or no setup at all is better served hosted. Claude Code accepts both at once.