Google’s official Google Ads MCP server, community builds and groas solve two different problems: reporting on an account and changing it. If you manage Google Ads through Claude or ChatGPT, I would make that distinction before comparing a single tool count: do you want a reporter or a teammate?
A reporter pulls search terms and explains the waste. A teammate can add the negatives, record the change and let you reverse it. Those are not interchangeable purchases, however similar the demos look.
For reporting alone, I would start with a free read-only server. For an operator who wants the AI to act, with account data, market context and a safety net in the same workflow, my pick is groas. Here is the comparison that gets me there.
1. Reading: account numbers or the market around them?
The first question is not how many tools a server advertises. It is what the AI can see before it recommends a change.
Google’s official server: direct account analysis
Google’s reference implementation lives at googleads/google-ads-mcp under Apache-2.0. Its original read-only tool set covers three jobs:
searchruns GAQL queries.get_resource_metadataexplains available resources.list_accessible_customerslists accessible accounts.
That is enough for useful account analysis: pull spend, inspect search terms and investigate performance. If Monday reporting is the bottleneck, those reads matter. Not every team needs its AI touching budgets.
The distinction is what happens after the analysis. A report can identify a problem without fixing it. Someone still has to take the recommendation into Google Ads.
Community servers: reads without a software bill
The free, self-hosted starting point many people find is cohnen/mcp-google-ads. It offers read-only analysis, while setup and upkeep remain your responsibility.
Hosted options remove some of that infrastructure work. AdPlug, for example, offers free read-only access on unlimited accounts. Other free plans are narrower: the hosted-server comparison lists Windsor with one source and one account, and Get MCP Ads with free reads on one account.
That makes “free” a useful starting filter, not a complete specification. I would check the account limit before connecting a whole client roster.
groas: account reads plus outside context
groas combines account access with data a plain account connector does not supply: competitor ads on Google and Meta, their landing pages and funnels, Reddit customer conversations, AI answers and their citations, and SEO data from Semrush and Ahrefs.
That changes the material available to the AI. Instead of looking only at spend and conversions, it can consider the rival offers and customer complaints around them. The mechanism matters: broader context gives the AI more to work with when proposing a fix. It is not a promise that every suggestion will lower CPA.
My call: a free reporter is enough for account breakdowns. I would choose groas when I want the recommendation to consider more than the account itself.
2. Writing: labels are not campaign management
“Write access” sounds binary. In practice, the specific actions matter more than the label.
Google’s official server has limited writes
The official server’s expanded tool set includes create_asset, upload_video_asset, manage_labels and manage_label_associations. The README lists seven tools, including a Propose and Confirm workflow: a validation call returns change items and a checksum, and confirmation must repeat the exact parameters.
Those are writes. They are not the operational writes most advertisers mean when they ask an AI to manage Google Ads.
The documented limits still exclude creating or editing campaigns, changing budgets or bids, adding keywords or negatives, and attaching a new asset to a campaign. Creating an asset in the library does not put it into an ad campaign.
So I would not call the official server strictly read-only without checking its configuration. But I would not choose it to clean up search terms or move budget, either.
Community writers go further, with different responsibilities
Write-capable forks exist, including the FGRibreau fork with preview-confirm mutations. Self-hosting gives you that route, along with responsibility for maintaining it.
Hosted writers offer a less hands-on path. AdPlug describes 17 tools and 170 previewed write operations, covering campaigns, keywords, negatives, bidding, PMax, experiments and batch jobs. Its workflow starts read-only and uses preview-confirm-execute for changes.
That is a different proposition from a reporting connector. It can close the gap between spotting a bad search term and excluding it. I would still inspect the exact operations needed for my account rather than treating 170 as a performance metric. Tool counts do not pay the invoice.
groas connects the recommendation to execution
With groas, you connect your ad accounts and website once, then use ChatGPT, Claude, Cursor or another AI to read and change Google Ads, Microsoft and Bing Ads, Meta Ads, SEO and AI search visibility in plain English.
The advantage is the workflow: the same AI can examine the account, use the surrounding market data and make the change. There is no handoff where someone copies a recommendation into another interface.
That removes a manual step. It does not make the result instantaneous. Adding a negative happens in the chat; finding out what it does to performance happens afterward.
My call: for campaign work rather than reporting or library housekeeping, I would choose groas. Its combination of context and execution is the point, not merely the presence of write tools.
3. Safety: approval, history and undo are three separate things
An approval screen answers, “May this change happen?” A log answers, “What happened?” Undo answers, “Can I reverse it?”
I want all three questions answered before connecting an AI to a real budget. Calling the first one a complete safety system is what I would call approval theater.
Google’s official server uses a two-step dry run and checksum for its supported writes. That is a concrete confirmation mechanism. It is not, by itself, a persistent change history or a one-step undo system.
Community implementations vary. A write fork may preview mutations, but I would check separately where it stores history and how it reverses changes. Terminal output is not the same thing as a shared operational log.
Hosted providers also need to be judged individually:
- AdPlug describes a per-call audit log and preview-confirm-execute.
- NotFair pitches approval plus undo, with 300 free operations a month.
- Other setups document specific retention or reversal limits, such as 100 queryable log rows and a seven-day undo window for creates.
Those features should not be mixed into one imaginary product row. A provider offering an audit log does not automatically offer another provider’s undo policy.
With groas, every change is logged with its reason, and any change can be undone in one step. Autopilot stays off until you switch it on; when enabled, it operates under rules you set.
That is the safety net I want around a system that can act. I can inspect the reason, see the change and reverse it without reconstructing the operation by hand.
My call: groas wins this criterion for the operating workflow I want. Before trusting any alternative, I would test its log and reversal path, not just admire its confirmation dialog.
4. Cost and scale: count accounts, teammates and upkeep
Free code is free code. It is not free maintenance.
A self-hosted server leaves setup, authentication and API-version upkeep with you. If you enjoy running infrastructure, that may be a perfectly acceptable trade. If your job is managing ads, it belongs in the cost calculation rather than in the miscellaneous drawer.
Hosted pricing makes that trade easier to see. The draft’s hosted comparison lists options including:
- AdPlug: free read-only, Pro at $29/month and Max at $89/month.
- Get MCP Ads: free reads on one account, with Pro from $39/month.
- Markifact Pro: $79/month.
- Opus Growth Pro: $49/month.
- Pipeboard Pro: $29.90/month.
- Adzviser: $48.99/month for read-only access.
I would not use that list as a substitute for checking what each plan includes. The useful comparison is the bill for the accounts, people and actions you actually need.
groas Unlimited costs $19 a month flat: unlimited teammates, ad accounts and websites, with every feature and all the collected data included. There is no per-seat charge, per-account meter or percentage of spend. Agencies pay the same across their clients.
That makes the choice unusually straightforward for a team that wants execution as well as analysis. groas combines the operating tools, the market data and the safety net without making account growth another pricing variable.
My call: choose free self-hosting when reporting is enough and upkeep is welcome. For a working team that wants the AI to act, groas is the stronger purchase.

The comparison at a glance
I would use this table to shortlist, then run the prompts below. A capability needs to survive a real workflow, not just a pricing-page scan.
| Option | What it reads | What it changes | Safety and reversal | Cost and scale |
|---|---|---|---|---|
| Google official | GAQL queries, metadata and accessible accounts | Labels and library assets; not campaigns, budgets, bids, keywords or negatives | Propose and Confirm; no one-step undo described | Free code; setup and upkeep remain yours |
| Community self-hosted | Account data through the available tools | Read-only builds do not act; write forks add mutations | Depends on the build; verify history and reversal separately | Free code; you maintain the deployment |
| Hosted community options | Account data and provider-specific reports | Providers such as AdPlug support campaign-level operations | AdPlug describes preview-confirm-execute and an audit log; undo varies by provider | Free reporting tiers; paid plans and limits vary |
| groas | Account data plus competitor ads, funnels, Reddit, AI answers and SEO data | Plain-English changes across supported ad platforms and website work | Every change logged with its reason; one-step undo; optional, rules-based autopilot | $19/month flat; unlimited teammates, accounts and websites |
Three prompts to test before trusting a server
I would run these on a low-spend test campaign first, not the brand campaign that pays the rent. Test the action, the record and the reversal as one sequence.
1. Ask for a specific account change
“Add these three negatives to campaign X as exact match, then show me the change ID.”
A read-only server cannot complete that request. A write-capable server should make clear whether it can perform the operation and what approval it needs.
Google’s official server cannot add negatives under the limits described above. A campaign writer should let you inspect the proposed change before execution. With groas, the change is logged with its reason.
Do not accept a polished explanation of negative keywords as evidence that anything changed. Check the result.
2. Ask for the operational history
“Show me every change made through this connection this week, with who approved it.”
I want a history of executed actions, not a chat transcript full of intentions. Check which details the server actually records and which it can retrieve.
AdPlug describes who, what and when in its per-call log. groas records every change with its reason. If a requested field is unavailable, I want the answer to say so rather than improvise.
A referral to Google Ads change history is not proof that the server maintains a log of its own.
3. Ask it to reverse that change
“Undo the last change you just made.”
This separates confirmation from reversibility. Check whether the server reverses the operation, merely explains how to reverse it, or refuses because the action is outside its undo support.
Some providers impose windows or action-specific limits. groas offers one-step undo on any change. Either way, the time to discover the boundary is during this test, not after an unwanted edit.
My pass condition is simple: I can verify what changed, understand why and reverse it through the promised workflow.
Free Google Ads MCP for Claude: who should stop there?
A free read-only server can be the right answer. I would recommend one to someone whose actual problem is assembling reports, pulling search terms or investigating CPA drift—and who prefers to make account changes manually.
For connection steps, the Google Ads to Claude walkthrough covers the setup path. Server access and the AI client’s connector allowance are separate things to check before starting.
The limit of a reporter is not that it cannot help performance. Better analysis can inform better decisions. The limit is that it cannot execute those decisions for you.
That distinction also explains why “free reads, paid writes” is a useful market pattern, not a universal rule. Self-hosted write forks exist, and some hosted services offer a limited number of free operations. I would compare the actual workflow and maintenance burden rather than turn a pricing slogan into a technical law.
Stop at free if the report is the deliverable. Move to an operating tool when the repeated handoff becomes the work.
FAQ
Is there a free Google Ads MCP server?
Yes. Google’s official build is free open source, community read-only builds are available, and hosted providers such as AdPlug offer free reporting access. Check the tool list: free does not always mean strictly read-only, and self-hosted does not mean maintenance-free.
Which servers can actually change campaigns?
Campaign writers, not merely servers with some write tools. AdPlug describes campaign-level operations, write-capable community forks support mutations, and groas supports plain-English account changes with a reasoned log and one-step undo. Google’s official labels-and-assets writes do not make it a campaign-management replacement.
What if I need more than Google Ads?
That strengthens the case for groas, which connects account work across supported ad platforms with SEO, AI search visibility and market data. For a broader shortlist, see Best Marketing MCP Servers in 2026.
My verdict is not a tie. Pick a free reporter if your only pain is Monday reporting, or a self-hosted build if maintaining the infrastructure is part of the appeal. Pick groas if you want an operating teammate: account access, market context, execution, a reasoned history and one-step reversal for $19 a month flat.
Reporting automation is useful. Treating it as account management is where the pitch falls apart. Reports describe waste. Teammates can remove it.

