How to Manage Google Ads With Claude

Published August 4, 2026 · 7 min read

Claude is the easiest of the major assistants to point at a live ad account, because MCP is native to it — connectors are a first-class feature rather than a workaround. This is the full path: connect, verify, audit read-only, then turn on writes with guardrails.

Why Claude specifically

Anthropic published the Model Context Protocol, and Claude Desktop, Claude Code and the Claude web app all support remote MCP servers directly. Practically that means: no plugin marketplace, no scraping, no pasting CSV exports. You add a URL, sign in with Google, and the ad account's tools appear in the conversation.

The second reason is tool-use behaviour. Running an account well means chaining calls — pull search terms, then pull the campaigns those terms sit in, then check conversion actions before recommending anything. Claude sequences that reasonably and asks before doing something destructive, which matters when the tools spend money.

Step 1 — Add the connector

Claude Desktop

Settings → Connectors → Add custom connector, then paste:

https://klienta.co/mcp

Claude Code

claude mcp add --transport http klienta https://klienta.co/mcp

Both use the same remote server over Streamable HTTP with OAuth 2.0. Nothing is installed locally and there is no Google Ads developer token to apply for. Screenshots for each client are on the setup page.

Step 2 — Authorize and verify

Claude opens Google's sign-in page; approve access for the account you want to manage. Then verify with the smallest possible request:

List my Google Ads accounts.

If customer IDs come back, you are connected. If nothing comes back, the connector is added but not authorized — reconnect from the same settings panel.

Step 3 — Spend a week read-only

Resist the urge to let it change things on day one. Read-only work is where most of the value is anyway, and it teaches you how the model behaves on your account. Prompts worth running:

• "Summarize the account: campaigns, budgets, bid strategies, conversion actions. Flag anything obviously misconfigured."
• "Search terms, last 30 days, 10+ clicks and zero conversions. Group by theme with total wasted spend per theme. Do not apply anything."
• "Month over month by campaign: cost, conversions, CPA. Flag moves over 25% and correlate with change history."
• "Which conversion actions have not recorded in 30 days while campaigns still optimize toward them?"
• "Show quality score distribution weighted by spend — where is the expensive low-QS traffic?"

Step 4 — Turn on writes, carefully

When you do allow changes, the pattern that works is propose → review → apply → keep the undo handy:

1. Set guardrails first. Budget ceilings, maximum percentage change, cap on entities per operation. These bound the worst case regardless of what the model decides.
2. Ask for the diff. "Show me exactly what you would change and the projected impact, then wait." Read it. Models state wrong numbers in exactly the same confident tone as right ones.
3. Apply in batches with a theme. One session for negatives, one for budgets. Mixed batches are hard to review and harder to reverse.
4. Keep the changeset ID. "Undo the last changeset" should be one sentence, not an archaeology project.

Step 5 — Make it a routine

The value compounds when it stops being a novelty. A workable weekly rhythm:

In Claude Code, these become project commands you re-run rather than retype.

What Claude will not do for you

• Know your business. It sees a 3× CPA, not the loss-leader strategy behind it.
• Beat Smart Bidding at its own game. It changes settings; the auction is still the auction.
• Be right because it sounds right. Verify numbers that will drive a real decision.
• Watch the account overnight. Chat is pull, not push. Scheduled monitoring is a different tool.

Also works with Meta

The same connector covers Meta Ads — 19 tools for campaigns, ad sets, ads, insights and delivery diagnostics — so cross-channel questions work in one conversation. See the Meta Ads MCP guide, and the ChatGPT version or the Gemini version if Claude is not your client.