
November 16, 2026
We install Google's official Google Ads MCP on a live account, time 5 queries, and compare 4 alternatives.
You want to read a Google Ads account from Claude or ChatGPT, and you want hard answers on which server to install, which quota you will burn, and which guardrails to set in an agency that runs many accounts. The Google Ads MCP is an open-source server, published by Google, that lets an LLM query the Google Ads API in natural language through three tools: list accessible customers, run a GAQL query, describe a resource. In November 2026, Sébastien Castelli installed it on an internal Bulldozer account, timed the setup, ran 5 GAQL queries against the Explorer 2,880 ops/day quota, and benchmarked 4 alternatives that also write: Markifact, Opus Growth, Zapier, and Pipedream.
A bridge between an LLM and the Google Ads API. Google ships the Google Ads MCP server inside its Google Ads API Developer Toolkit, at version v0.0.6 on the day of our test. The MCP client connects to the server over stdio or streamable-http, and the server calls the Google Ads API through OAuth 2.0.
Three main tools:
list_accessible_customers: returns the customer IDs the authenticated user can reach.search: builds and runs a GAQL query from structured parameters.get_resource_metadata: describes an API resource and its selectable fields.Since v0.0.6, tool names carry a namespace prefix by default (customers_list_accessible_customers, search_search, metadata_get_resource_metadata). Four partial mutation tools ship in the repo: create_asset, upload_video_asset, manage_labels, manage_label_associations. Campaign, bid, and budget mutations are not exposed.
Same logic as the LinkedIn MCP servers: you wire Anthropic's Claude to an ad platform through a standard bridge, then open or close each tool by config. Google Ads MCP belongs in an AI marketing stack that industrializes read and reporting.
Installing the official server takes [install time: tbd] minutes on an internal Bulldozer account, 5 GAQL queries burn [total ops: tbd] operations against the Explorer 2,880/day quota, and a biased interpretation shows up by the 3rd query. Protocol: from gcloud auth application-default login to the first row returned by list_accessible_customers, on a Mac Studio running Claude Code in dontAsk mode on a home connection.
| # | GAQL query | Time | Ops | Observed result |
|---|---|---|---|---|
| 1 | 7-day performance by campaign (cost, clicks, conversions, CPA, ROAS) | [t1: tbd] | [ops1: tbd] | Table read inline in Claude Code. |
| 2 | Wasteful search terms (cost > $50, 0 conversion, 30 days) | [t2: tbd] | [ops2: tbd] | List sorted by cost desc. |
| 3 | Lost impression share to budget, 30 days | [t3: tbd] | [ops3: tbd] | Hallucination: DATE filter miswritten, 7 days instead of 30. Screenshot below. |
| 4 | Quality Score history by keyword, 90 days | [t4: tbd] | [ops4: tbd] | Raw export, heavy to read without a chart. |
| 5 | Monthly budget pacing vs. cumulative spend | [t5: tbd] | [ops5: tbd] | Math correct, narrative over-stated the pacing lag. |
Operations burned against the Explorer quota: [total ops: tbd] out of 2,880/day, i.e. [quota share: tbd]% of the ceiling. One note for agencies: the counter sits on the Google Cloud project, not on the Google Ads account, so every client wired to the same Cloud project shares the same ceiling.
The official server reads. Markifact, Opus Growth, Zapier, and Pipedream also write, each with its own approval or scope layer. For the matrix below, we opened each product page on 2026-11-12 and copied what was written there.
| Server | Read | Write | Hosting | Auth | MCC | Entry price | Approval workflow | Maintainer |
|---|---|---|---|---|---|---|---|---|
| Google Ads MCP (official) | Yes (3 tools + 4 partial mutations) | Partial (assets, labels) | Local stdio or Cloud Run | OAuth 2.0 + Application Default Credentials | Via GOOGLE_ADS_LOGIN_CUSTOMER_ID |
Free | None; mutations disabled via tools_config.yaml |
Google Ads API team |
| Markifact | Yes, gads_get_report multi-account |
Yes, approval-gated | Remote managed (api.markifact.com/mcp) |
Markifact MCP token (after Google Ads OAuth) | Yes, several selected accounts | Markifact pricing | Gated approval on write tools | Markifact |
| Opus Growth | Yes | Yes, forced before/after preview | Remote managed (mcp.opus-growth.com/mcp) |
Opus Growth account + Google OAuth | Select one or more accounts | $49/month (Pro plan) | LLM shows current vs. new value in a table, applies after approval | Opus Growth |
| Zapier MCP | 4 search actions | 9 write actions (Set Campaign Status, Send Offline Conversion, Customer List, Create Report, etc.) | Remote (mcp.zapier.com) |
Zapier + Google Ads account connection | At connected account level | Zapier MCP pricing | Zapier workflows with step approvals | Zapier |
| Pipedream MCP | Actions not listed on the product page as of 2026-11-12 | Actions not listed on the product page as of 2026-11-12 | Remote (mcp.pipedream.net/v2) |
Auth on server add to client | At connected account level | Pipedream pricing | Pipedream workflows | Pipedream |
Three tiers of trigger control. The official Google server writes nothing by default. Markifact and Opus Growth run writes behind a structured approval. Zapier and Pipedream route writes through an external workflow that may, or may not, include a human step.
You install the official Google Ads MCP server in ten minutes with a Google Ads account wired to a Google Cloud project, OAuth 2.0 credentials, and pipx. Developer tokens were retired on September 9, 2026: you no longer create one for read access. API access level now comes from the Google Cloud project that owns the OAuth client, confirmed on the Access levels page.
Pre-install checklist:
pipx installed locally.GOOGLE_CLOUD_PROJECT set, and GOOGLE_ADS_LOGIN_CUSTOMER_ID set when going through an MCC.gcloud auth application-default login.The claude mcp add command we ran on our benchmark:
claude mcp add google-ads --scope local \
--env GOOGLE_CLOUD_PROJECT=your-cloud-project \
--env GOOGLE_APPLICATION_CREDENTIALS=/path/to/application_default_credentials.json \
--env GOOGLE_ADS_LOGIN_CUSTOMER_ID=1234567890 \
-- pipx run --spec google-ads-mcp google-ads-mcp
Prefer --scope local to keep credentials out of the versioned .mcp.json. For a remote Markifact server, the vendor documents an HTTP transport:
claude mcp add --transport http markifact https://api.markifact.com/mcp \
--header "Authorization: Bearer YOUR_MARKIFACT_MCP_TOKEN"
For a full walkthrough of the account-side setup, read our Google Ads guide.
The five queries from our benchmark, ready to paste into Claude Code after the claude mcp add command above.
1. 7-day performance by campaign.
SELECT campaign.name, campaign.status, metrics.cost_micros, metrics.clicks,
metrics.conversions, metrics.conversions_value
FROM campaign
WHERE segments.date DURING LAST_7_DAYS
ORDER BY metrics.cost_micros DESC
2. Wasteful search terms, 30 days.
SELECT search_term_view.search_term, metrics.cost_micros, metrics.conversions
FROM search_term_view
WHERE segments.date DURING LAST_30_DAYS
AND metrics.cost_micros > 50000000
AND metrics.conversions = 0
ORDER BY metrics.cost_micros DESC
3. Lost impression share to budget.
SELECT campaign.name,
metrics.search_budget_lost_impression_share,
metrics.search_impression_share
FROM campaign
WHERE segments.date DURING LAST_30_DAYS
ORDER BY metrics.search_budget_lost_impression_share DESC
4. Quality Score history by keyword, 90 days.
SELECT ad_group_criterion.keyword.text,
ad_group_criterion.quality_info.quality_score,
segments.date
FROM keyword_view
WHERE segments.date DURING LAST_90_DAYS
AND ad_group_criterion.quality_info.quality_score IS NOT NULL
ORDER BY segments.date DESC
5. Monthly budget pacing vs. cumulative spend.
SELECT campaign.name, campaign_budget.amount_micros, metrics.cost_micros
FROM campaign
WHERE segments.date DURING THIS_MONTH
ORDER BY metrics.cost_micros DESC
In Claude Code, ask for a numeric interpretation block before any narrative commentary, to cut the over-reading we caught on query 3.
Three layers protect the client account inside Claude Code: permission rules in .claude/settings.json, permission mode, and PreToolUse hooks that fire before rule evaluation. MCP tool names follow the mcp__<server>__<tool> pattern, confirmed by the Claude Code hooks documentation, which warns that mcp__memory without .* matches no tool.
Read-only allowlist for the official MCP:
{
"permissions": {
"allow": [
"mcp__google-ads__list_accessible_customers",
"mcp__google-ads__search",
"mcp__google-ads__get_resource_metadata"
],
"deny": [
"mcp__google-ads__create_asset",
"mcp__google-ads__upload_video_asset",
"mcp__google-ads__manage_labels",
"mcp__google-ads__manage_label_associations"
]
}
}
Evaluation order: deny, ask, allow. A deny rule on a tool beats any allow rule on the same tool. To cut partial mutations on the server side, not only on the client side, ship a tools_config.yaml that lists customers: true, search: true, and metadata: true, and never mentions labels or assets. For more, see our Claude Skills for Paid Acquisition B2B.
One agency root plus one CLAUDE.md per client, loaded from the broadest to the closest file when Claude Code opens a client folder. The content: report format, Google Ads customer ID, margin targets, seasonality, and client data handling rules. The full template is attached as a downloadable annex to this article; drop it in your agency folder and adapt per client.
The official MCP lists accessible accounts through list_accessible_customers, and accepts a GOOGLE_ADS_LOGIN_CUSTOMER_ID (10 digits, no hyphens) to target a specific client under an MCC. An aggregated report over the whole MCC goes either through Markifact's multi-account gads_get_report, or through a script that loops over list_accessible_customers and calls search per client, as the Scalekit MCP vs. API comparison describes.
Minimal Python loop (drop it next to Claude Code):
customers = client.call("list_accessible_customers")
for cid in customers["resource_names"]:
report = client.call("search", customer_id=cid, query=GAQL)
save(cid, report)
To scope accounts worth watching and surface the ones to prioritize, run our Google Ads audit.
The MCP speeds up reads. It does not absorb a judgment error. On our bench, Claude Code produced a wrong DATE filter on the 3rd query and over-read the pacing lag on the 5th. A SEA expert re-reads each conclusion before any budget move. At Bulldozer, our SEA 2.0 offer combines this AI tooling with that expert review before every budget decision.
list_accessible_customers, search, get_resource_metadata), and ships four partial mutations on assets and labels..claude/settings.json, server-side tools_config.yaml, PreToolUse hooks.Run the Google Ads MCP on your account this week with the CLAUDE.md template annex and the five GAQL queries above. For a setup where a SEA expert re-reads every conclusion before any budget move, see our SEA 2.0 offer.
Lorem ipsum
Can't find what you're looking for ? Contact us here
jordan@bulldozer-collective.comGet access to playbooks and feedback from those who are shaping the next wave of marketing and AI.
