Free self-hosted Google Ads MCP servers do one thing well: they let your AI client read your account — GAQL queries, account lists, performance tables — for the price of running them yourself. The best-known servers in the category expose three to five tools, all read-only. AdCopilot is a hosted connector with 54 read-and-write Google Ads tools (as of v2.16.0) and a deliberate limit: it cannot delete anything. This page lays out both options without pretending either is wrong.
What the free self-hosted servers actually are
An MCP server is the bridge between an AI client — Claude, ChatGPT, Cursor — and the Google Ads API. The free self-hosted category is a set of open-source repos you clone, configure and run on your own machine or cloud. The best-known server in the category exposes exactly three read-only tools: a GAQL search, a resource-metadata lookup and an accessible-accounts list. The GitHub favourite, cohnen/mcp-google-ads, adds a couple more read tools — account listing, GAQL execution, campaign and ad performance — and is the most-linked community option in every registry. A newer entrant, TrueClicks' google-ads-mcp-js, removes the developer-token hurdle by authenticating through the vendor's own service, still read-only.
Common shape across the category: free licence, public code, read-only tools, and you run it.
What they are genuinely good at
Credit first, because it is real:
- Zero licence cost. The only bills are Google's API access and your time.
- Inspectable code. You can read every line before it touches your account. No hosted product offers that.
- A direct data path. Your account data flows from Google to your machine to your AI client. No third party sits in the middle, because you are the third party.
- Read-only by construction. A server with no mutate tools cannot spend your money no matter what the model does. For a cautious first experiment with an AI agent on Google Ads, that is a legitimate safety posture — the same capability-absence logic AdCopilot applies to deletion.
If your goal is "ask Claude questions about my account" and nothing more, a free self-hosted server answers it.
What self-hosting actually requires
The licence is free; the setup is not. Before the first query runs you need:
- A Google Ads developer token — applied for through your own manager account, with an approval wait. What that involves is covered in developer tokens and API access.
- Your own OAuth client — a Google Cloud project, consent screen and credentials, wired into a local config file.
- A runtime — Python or Node on your laptop, or a small cloud service if you want it available from more than one machine or for more than one person.
- Upkeep — the Google Ads API versions roughly three times a year, and a community repo only tracks it if its maintainer does. The most-linked community server's last push, at the time of writing, was October 2025. The code still runs; nobody is on the hook when it stops.
None of this is hard for a developer. All of it is real work, and it repeats per teammate — there is no team model, no per-member sign-in, no audit trail unless you build them.
What AdCopilot adds — and what it refuses
AdCopilot is the same protocol with the opposite operating model. You add the connector URL to any MCP client — Claude, ChatGPT, Copilot, Gemini CLI, Cursor, OpenCode — and sign in with your own Google account. No developer token, no OAuth client of your own, no runtime.
The capability difference is the point. AdCopilot exposes 54 Google Ads tools (as of v2.16.0), read and write: create search and Performance Max campaigns, ad groups, responsive search ads, keywords and negatives, ten asset types, geo targeting, ad schedules, device bid adjustments, budget changes, pause and enable. Your AI client asks approval before each write by default; reads change nothing and can be allowed once.
And one capability is deliberately absent: it cannot delete anything. The four remove tools are never exposed to hosted connectors, and any mutate carrying status REMOVED is refused server-side — in any letter case. That is enforcement in the server, not a prompt asking the model to behave. The full model is on the security page.
The comparison, in one table
| Capability | Free self-hosted servers | AdCopilot |
|---|---|---|
| Licence cost | Free (open source) | 7-day free trial for a new workspace, then Free plan |
| Tools exposed | 3–5, read-only | 54 Google Ads tools (as of v2.16.0), read and write |
| Campaign creation | No | Yes — search and Performance Max, arriving paused |
| Deletion | No tools to delete | Cannot delete — remove tools absent, REMOVED refused server-side |
| Developer token | You apply and wait | Not needed (bring-your-own supported) |
| OAuth setup | Your own Cloud project | Sign in with Google, done |
| Hosting and updates | Yours | Hosted, maintained |
| Team access | One config per person, DIY | Per-member Google sign-in, individually revocable |
| Audit trail | Build your own | Every call logged — tool, account, outcome, timestamp |
| Support | Repo issues | A vendor accountable to you |
When a free self-hosted server is the right choice
Choose free and self-hosted when reporting is the whole job and you have the hands to run it: you are technical, you already hold a developer token or do not mind the application, your data-path policy prefers no intermediary, and nobody needs the agent to change anything. It is also the right way to learn what a Google Ads MCP server even is before spending money — run one for an afternoon and the category stops being abstract.
When a hosted write-capable connector is
Choose hosted when the point is management, not just measurement: you want the agent to add the negatives it found, build the campaign it drafted and move the budget it argued for — with an approval on every write. Choose it when more than one person needs access with their own identity, when someone must be able to answer "what changed and who approved it", and when nobody on the team wants to own Python environments or API version bumps.
The sane hybrid: run both
These options are not exclusive. A pattern that works: keep a free read-only server on a developer's machine for ad-hoc GAQL spelunking, and use a hosted connector for the operational loop — audits that end in fixes, weekly routines, campaign builds. The free server costs nothing to keep around; the hosted one is the only one of the pair that can act.
If the acting half is what you are missing, the trial is the cheap test: start free — seven days on the full Pro plan for a new workspace, no card — and run the same questions through both. The difference you feel on day two is the difference this page just described.