September 2, 2026 · 8 min read
Buyer's guide

Google Ads MCP vs Agent Planners: when a read-only MCP server is enough — and when it isn't

TL;DR

Google open-sourced an official Google Ads MCP server in early 2026: two tools — list accessible customers, and a GAQL-powered search — that let Claude or ChatGPT query your account in plain language. It's genuinely good, genuinely free, and deliberately read-only. That last word is the whole comparison: it can tell you what's wrong and cannot do anything about it, and a chat session can't run at 9am on Monday.

What the official MCP server actually is

  • Open-source and free, published by Google's marketing-solutions team.
  • Two tools: listing the customer accounts you can access, and a GAQL-powered search.
  • Strictly read-only — it cannot modify bids, pause campaigns or create assets. That's a deliberate containment decision, not an oversight.
  • You host it and bring credentials — it wants a developer token and OAuth setup most account managers have never generated.
Hosted third-party alternatives (Adzviser, Porter, Blend, AdKit and others) remove the setup work and are often near-free. If your question is "can I ask Claude about my Google Ads data", the honest answer is yes, cheaply, today.

Where it's the right tool and we're overkill

For ad-hoc diagnostics — "what did I spend last week", "which campaigns have zero conversions", "pull the search terms for this ad group" — an MCP server in your existing assistant is faster to reach for and costs nothing. If that's the whole job, use it. We publish our own MCP endpoint and support connecting your own custom MCP servers precisely because the protocol is good.

What a chat session structurally cannot do

NeedRead-only MCP in chatAgent harness
Answer a question about the accountYes — this is what it's forYes
Actually apply the fixNo — read-only by designYes, with mandatory approval on each write
Run at 9am Monday without youNo — a chat needs a human to start itScheduled, with results emailed
Remember your rules between sessionsNo — each conversation starts blankPer-account memory: targets, protected campaigns, excluded geos
Fan one audit across 40 MCC sub-accountsOne at a time, by handOne goal, per-account dry-run
Join Ads cost with your own revenueNot without another connection and a place to computeWrites SQL against your warehouse
Produce a client-ready dashboard on a linkNoSelf-refreshing page you can revoke

The part people underestimate: memory

A chat session forgets. Every time you ask about the account you re-explain that brand campaigns are never to be paused, that your CPA target is $40, that you don't bid in certain regions. That re-briefing is fine once and corrosive on the fiftieth run — and it's the exact thing that makes unattended scheduling possible or impossible. Rules that live with the account get applied whether or not anyone remembers to state them.

The honest decision rule

  • Just want to ask questions? Use an MCP server. Free, two minutes, genuinely fine.
  • Want the work done, on a schedule, with a record? That's a different product category, and the read-only server can't grow into it — being read-only is its design, not its roadmap.
  • Want both? They coexist. Keep the MCP server in your assistant for ad-hoc questions and use a harness for the recurring, governed work.
For how a governed write path differs from an ungoverned one, see is it safe to give an AI agent access to your ad account.

Frequently asked questions

Is there an official Google Ads MCP server?
Yes — Google open-sourced one in early 2026, exposing account listing and a GAQL-powered search. It is strictly read-only and cannot modify bids, pause campaigns or create assets.
Can the Google Ads MCP server change my campaigns?
No. The official implementation is read-only by design, for diagnostics and analytics. If you need changes applied, you need something that also has a governed write path.
Why use Agent Planners if the MCP server is free?
For asking questions, you often shouldn't — the free server is fine. The difference is everything a chat session structurally can't do: apply changes under an approval gate, run on a schedule unattended, remember your account rules between runs, fan work across an MCC, and produce shareable dashboards.
Do you support MCP yourselves?
Yes — we expose our own MCP endpoint so external agents can drive tasks, and you can connect your own custom MCP servers as tools. This isn't an argument against the protocol; it's about what a single chat turn can and can't accomplish.
Can I use both an MCP server and an agent harness?
Yes, and many teams do — the MCP server in your assistant for ad-hoc questions, the harness for recurring governed work. They aren't mutually exclusive.
Run your ad ops with an agent you can trust

Start free — 2,500 credits a month, no credit card. Every write waits for your approval.

Keep reading