Resources/Comparisons

The best Google Ads MCP servers in 2026, compared honestly

Daniel Popa·Founder, AdPlug·September 2, 2026·13 min read

Updated September 3, 2026

There are now at least ten serious ways to connect Google Ads to Claude, ChatGPT, or Cursor over the Model Context Protocol, and they are wildly different products wearing the same label. Some are read-only reporting bridges. Some can create campaigns. One is Google's own. I run Google Ads accounts for a living and I build one of the tools on this list, so this comparison is opinionated but checkable: every claim is either from my own testing or from the vendor's published documentation, and I say which is which.

The short version

If you want the table first, here it is. The rest of the article explains each row and who each server is actually for.

My one-paragraph verdict: if you are a developer who only needs reporting, run Google's official server and pay nothing. If you are a PPC specialist or an agency and you want an AI that can both audit an account and safely change it, you want a hosted server with real write depth and a preview step, and that shortlist is AdPlug, NotFair, and Adspirer. I build AdPlug, so read the methodology below and weigh my bias accordingly.

What a Google Ads MCP server actually is

MCP, the Model Context Protocol, is the open standard that lets AI assistants call real tools. A Google Ads MCP server is a bridge that exposes the Google Ads API as tools an assistant can use: list accounts, pull a search-terms report, check budget pacing, and, on the write-enabled servers, add negative keywords, pause ads, or create campaigns. Instead of exporting CSVs into a chat, the assistant queries the live account and acts on it.

The servers split along three lines that matter more than any feature list:

• Hosted vs self-hosted. Self-hosted servers are free but require your own Google Ads API developer token, OAuth credentials, and somewhere to run the code. Hosted servers give you a URL, you sign in with Google, and you are connected in minutes. Hosted is also the only practical route to ChatGPT, which expects remote connectors.
• Read-only vs write-enabled. Most integrations quietly stop at reporting. Whether the server can actually add the negatives it just recommended is the single biggest capability gap between the options below.
• Safety model. For write-enabled servers, what stands between the model and your live budget: a preview step, an approval gate, an audit log, or nothing.

How I compared them

I checked every server in early September 2026 against the same questions: hosted or self-hosted, read or write access, how deep the Google Ads write surface goes, what the safety model is for changes, which other ad platforms the same vendor covers, what the free tier really includes, and whether it connects to Claude, ChatGPT, and Cursor. Capability and pricing claims for competing products come from their public sites and docs, linked in each section, because I do not have production accounts on all of them. Where a vendor does not publish a number, I say so instead of guessing.

AdPlug: hosted, deep writes, preview on everything

AdPlug is a hosted MCP built specifically for people who run paid media all day. On Google Ads it exposes 16 tools covering account discovery across MCC trees, campaign and conversion performance, search-term audits, keyword research, raw GAQL, and a write pipeline with 148 distinct operations: campaigns, ad groups, ads, keywords and negatives, bidding strategies, audiences and remarketing lists, Performance Max asset groups, Demand Gen, shared sets, labels, experiments, and bulk batch mutations.

The safety model is the part I refuse to compromise on, because my own client accounts are connected to it. Every write goes through a preview tool that shows the exact operation before a separate execute step runs it, accounts are read-only by default, and every tool call lands in an audit log with who, what, and when.

Pricing is built to be the generous row in the table above, on the axis agencies feel first: every plan, including Free, allows unlimited ad accounts, so you never pay per connected client. Free is read-only forever on one platform of your choice, with 500 requests in your first month and 50 a month after, enough for a real trial audit. Pro at $29 a month covers all platforms with 5,000 requests and 200 previewed edits monthly, and Max at $89 removes the meters entirely: unlimited requests, unlimited edits, still unlimited accounts. Most hosted rivals in this comparison meter every plan by operations. The full breakdown is on the pricing page.

The same connector also covers LinkedIn Ads and Microsoft Ads, so a weekly check across search and B2B social happens in one conversation. It works with Claude (including the connector directory), ChatGPT, Cursor, and Claude Code, and the setup guide takes about two minutes. The honest limitation: AdPlug is built for marketers operating accounts, not for developers embedding ads data into their own agents, and there is no self-hosted option.

Google's official MCP server: first-party and read-only

Google shipped an open-source Google Ads MCP server in October 2025, with a developer integration guide that covers running it locally or deploying it on Google Cloud. It is the most authoritative answer to "can my AI query Google Ads" and the least finished answer to everything after that. The server is deliberately small, focused on executing GAQL queries and retrieving campaign data, and it is read-only: no negatives, no budget changes, no campaign creation.

The bigger filter in practice is setup. You need your own Google Ads API developer token, an OAuth client, and the comfort to run a Python server and keep credentials safe. For a developer or a data team that already has API access, that is an afternoon and the price is zero. For a working marketer it is usually the point where the project stalls. If your need is reporting and analysis only, and you clear the technical bar, this is the option I would pick, because first-party GAQL access with no middleman is hard to argue with.

The open-source route: cohnen/mcp-google-ads and friends

Before Google's server existed, the community built its own. The most established is cohnen/mcp-google-ads, which connects Google Ads to Claude and Cursor for natural-language analysis of campaign data. Like Google's server it is free, self-hosted, and needs your own developer token and OAuth setup, and it is oriented at reading and analyzing rather than managing. Other community servers, like FGRibreau's write-capable one, add preview-confirm mutations on top.

The open-source route makes sense when you want zero vendor dependency, you are happy maintaining credentials yourself, and local-first matters to you. Its limits are the same as the official server, plus the maintenance is on you when the Google Ads API versions roll over, which they do twice a year, retiring old versions as they go.

NotFair: approval-gated writes with an undo

NotFair is a hosted Google Ads MCP whose pitch is accountable writes. Per its published comparison from August 2026, it exposes 119 Google Ads tools, gates changes behind an approval step with an undo afterwards, supports changing image and video assets, and gives new accounts 7 unlimited days followed by a free tier of 300 operations a month. The same company publishes servers for analytics, search, and CRM alongside ads.

The approval-and-undo lifecycle is a genuinely good idea, and NotFair documents its methodology in public, which I respect and have borrowed for this article. The trade-offs to check for your case: the operation caps are the pricing axis, so heavy automation runs into the meter, and the ads coverage is Google-centric, so a B2B team that also runs LinkedIn needs a second product.

Adspirer: write-enabled with tiered caps

Adspirer is a hosted MCP for Google and Meta with a preview-confirm workflow on writes. Its published materials describe 60+ Google Ads tools, image upload support, and YouTube video handling, with plans tiered by monthly operations from small caps up to thousands. It sits in the same category as AdPlug and NotFair: sign in, connect, and let the assistant both read and change accounts with a confirmation step in between.

If your spend is concentrated on Google plus Meta and your monthly change volume is predictable, it is a credible pick. As with NotFair, the operation caps are the thing to model before committing, especially if you plan bulk work like restructures or large negative-list pushes.

Markifact: MCP inside an automation suite

Markifact approaches the problem from the automation side: it is a marketing automation platform whose MCP server plugs Google Ads into Claude, ChatGPT, and other agents, with write access for campaign creation, budget management, and keyword optimization. Its published onboarding gives 200 one-time credits to try it, and the wider platform advertises 1,000+ operations across many marketing tools.

Choose it if you want scheduled, workflow-style automation and the MCP is one entry point among several. If you only want the live account-in-a-chat experience, the suite around it is more product than you need.

Quick takes: AdKit, Zapier, TrueClicks, Windsor.ai

AdKit

AdKitpositions directly against the official server's biggest gap: it offers full read and write access to Google Ads where Google's own MCP can only pull reports, and covers Meta as well. A reasonable hosted pick for operators, with a smaller public footprint than the others in the write-enabled group.

Zapier MCP

Zapier's MCPexposes its Google Ads actions to any MCP client. The strength is breadth: the same connector reaches thousands of apps, so one setup covers your CRM and your ad account. The weakness is that the actions are generic integration primitives, not a PPC workflow surface, and usage rides on Zapier's task pricing.

TrueClicks

TrueClicks, the PPC auditing company, offers a free MCP server covering Google, Microsoft, and Meta Ads. It is read-only, so think of it as a no-cost reporting bridge with an auditing pedigree rather than a management tool. A good first taste of ads-in-your-AI with zero commitment.

Windsor.ai

Windsor.ai comes from the marketing data world: its MCP exposes the same Google Ads data it pipes into warehouses and dashboards, alongside 300+ other sources. It is for analysts blending ads data with everything else, not for making changes.

The safety bar for agencies and enterprises

If the accounts you would connect belong to clients, the comparison changes character. A misfired budget change on your own account is a bad afternoon. On a client account it is a fired agency, so the questions that matter are not about tool counts, they are about control: who authorized the connection, what can it change, what did it change, and how fast can you shut it off. Before connecting any MCP server to client spend, make the vendor answer five things in writing:

• OAuth, never shared logins. Access should come from a scoped Google grant you can revoke from your Google account at any time, not from a password or a shared developer token you cannot audit.
• Read-only by default. Write access should be something you switch on deliberately, per account, not the starting state.
• A preview or approval step on every write. The exact operation, in plain terms, before it runs. No preview, no client accounts.
• A per-call audit trail. Who ran what, on which account, with which parameters, when. This is what protects you in the awkward client meeting.
• Credential handling and paperwork. Encrypted token storage, a DPA you can countersign, GDPR posture, and instant revocation from both sides.

This checklist is the reason AdPlug looks the way it does: OAuth-only access, read-only defaults, a preview-confirm-execute gate on all 148 write operations, an audit log entry for every tool call, Fernet-encrypted credentials, a signable DPA, and verification by both Google and the LinkedIn Marketing Developer Program, plus team seats on the agency tiers so the whole team works from one governed setup. I run my own client accounts through it, which is the strongest safety claim I can make. To be fair to the field: NotFair's approval-with-undo lifecycle and Adspirer's preview-confirm flow answer the same questions in their own way, and any of the three clears a bar most integrations never attempt. The servers to keep away from client accounts are the ones with write access and no preview layer at all.

How to choose

• You are a developer and only need reporting:run Google's official server. Free, first-party, full GAQL.
• You want zero vendor dependency and accept maintenance: cohnen/mcp-google-ads or another open-source server, self-hosted with your own credentials.
• You run accounts for a living and want safe write access without any setup: a hosted, preview-gated server. AdPlug if you run several platforms or many client accounts and want unlimited accounts on every plan; NotFair if the approval-and-undo lifecycle on Google-only work fits how you operate; Adspirer if Google plus Meta is your exact footprint.
• You mainly want ads data next to other business data: Windsor.ai for analysis, Zapier if the value is cross-app automation.
• You want to try the category for free in five minutes:TrueClicks read-only, or AdPlug's free tier if you want the audit workflows from the 30-minute audit on your own account.

Whichever you pick, the pattern underneath is the same and it is worth internalizing: the model matters less than the connection. An assistant with live account access and a preview gate will beat a smarter assistant staring at an exported CSV every week of the year. If you want to see what that feels like on your own spend, connect Google Ads or LinkedIn Adsand run one real audit. The read-only tier exists precisely so you can judge this category on your own account instead of anyone's comparison table, including mine.