Google publishes its own MCP server for Google Ads. It is read-only, it is good, and it runs on your machine. Whether that is enough depends on one question: is the person asking Claude about the campaigns you, or your team? Eight options, compared as they read in October 2026.
What changed in 2026
Two things make older guides misleading. First, Google retired the developer token on 9 September 2026: the API access level now belongs to the Google Cloud project behind the OAuth client, and the developer-token header is ignored. Any setup guide that starts with "apply for a developer token" is out of date, and so is any MCP server README that still requires one. Second, Google now maintains an official server, googleads/google-ads-mcp, documented on the Google Ads API developer site. That shifts the default: the question is no longer whether an Ads MCP server exists, but who runs it and for whom.
Google's official server
The official server has three tools: search for GAQL queries, get_resource_metadata to look up fields, and list_accessible_customers. Google's documentation is explicit that it is strictly read-only: no bid changes, no pausing, no creating. You install it with pipx and run it over stdio next to Claude Desktop, or deploy it yourself over Streamable HTTP, for example on Cloud Run. Authentication is an OAuth client, Application Default Credentials or a google-ads.yaml, plus a GOOGLE_PROJECT_ID whose project holds Explorer, Basic or Standard access. A hosted endpoint run by Google exists, but only for allowlisted pilot participants.
For one analyst on one laptop this is the right answer, and it costs nothing. The limits only show when more people want it: every one of them needs Python, a Cloud project or access to one, and a local config with a credential in it. In claude.ai in the browser or on a phone, a local stdio server does not exist at all. If you deploy it to Cloud Run instead, it runs under the one Google identity you deployed it with, and putting an authorization server in front of it is your job.
The eight options
- Google's official server. Read-only, three tools, local or self-deployed. Free.
- cohnen/mcp-google-ads, the best-known community server. Local Python process, OAuth client or service account, GAQL reads only. Its README still asks for a developer token, which no longer does anything.
- Composio, a developer platform for agent tools. 36 Google Ads tools, 25 of them writes, from mutating campaigns to batch operations. Managed OAuth app or your own. Free tier, then $29 a month plus usage.
- Windsor.ai, a marketing data platform with an MCP endpoint. Reads, and writes that include creating campaigns, ad groups and responsive search ads, setting budgets and bidding, and pausing or enabling. New items start paused; approval relies on the chat client asking before the call.
- Markifact, a hosted Google Ads MCP with a large set of operations, where every write requires approval. Pro $79, Team $249 a month according to its own blog.
- Adzviser, a hosted remote server, listed in Claude's directory, Google Ads read-only, made and hosted in the US.
- AnythingMCP, a German MCP gateway with 17 read-only Google Ads tools. You create your own OAuth client in Google Cloud and authorize it with one Google user. Cloud in Frankfurt or self-hosted under AGPL.
- Loopthink, our product, so weigh this row accordingly. One Google sign-in, no Cloud project of your own, read-only by default, every chat user with an identity and a role of their own, every call in an audit log under their name.
The comparison
| Google official | Community (cohnen) | Composio | Windsor.ai | AnythingMCP | Loopthink | |
|---|---|---|---|---|---|---|
| Works in claude.ai in the browser | Only if you deploy and secure it | No, local | Yes | Yes | Yes | Yes |
| Your own Google Cloud project needed | Yes | Yes | No, managed app optional | No | Yes, your own OAuth client | No |
| Write tools | None | None documented | 25 of 36 tools | Yes, including enable | None | Off by default; never enables |
| Per-person identity in the chat | No | No | Per connected account, built for developers | Workspace users | Roles admin, developer, viewer | Yes, roles per person |
| Audit log per tool call | No | No | Logs to inspect | Not described | Yes | Yes, under the person's name |
| Hosting | Your machine or project | Your machine | Not stated | Not stated | Frankfurt or self-hosted | Germany |
| Price | Free | Free | Free tier, then $29 plus usage | Not published on the MCP page | From €19 a month | Free for two connectors, team from €19 |
Sources: each vendor's own documentation and pricing page, linked above, as read on 17 October 2026. Where a page says nothing, the cell says so rather than guessing.
The question behind the table: who is Google talking to?
In every option here, Google sees one account. With the official server it is whoever ran gcloud auth; with a hosted service it is whoever pressed "Connect Google account". That is not a flaw anybody can fix, it is how Google Ads API access works for a shared tool. The difference is what sits between your colleague and that account.
If nothing sits there, everyone who can reach the server can do whatever that Google account can do, and Google's change history will show the account, not the colleague. For a read-only server that matters less: the worst case is a person seeing an account they should not see. For a server with write tools it is the whole question. A shared Admin login behind a chat can pause campaigns and move budgets, and six weeks later nobody can say who asked for it.
What helps is a layer that knows the person. Loopthink signs each chat user in with their own identity, decides per role which accounts and which tools they get, and writes every call into an audit log with their name, the query and, for changes, the before and after. Google still sees one account. Your audit log does not.
Which one fits
- You, alone, on your own machine: Google's official server. It is maintained by the people who build the API, it is read-only, and it is free.
- You are building an agent product: Composio. It is made for developers wiring tools into their own software, and its write coverage is the broadest here.
- You want the AI to run campaigns: look closely at how Windsor.ai and Markifact gate a write, and decide who in your team may approve one before you hand anyone an Admin login. The chat's confirm button knows nothing about your org chart.
- A team that should report, and maybe draft changes, without one shared login: that is the case Loopthink is built for, and the one where the identity row of the table decides.
Try it on your own account
Connecting Google Ads to Claude or ChatGPT through Loopthink takes one Google sign-in and no Cloud project. It is free for two connectors, with no card. Roles per person and field masking start at €19 a month.