Google, Meta, Microsoft and Amazon each ship an official ads MCP (Model Context Protocol) server; LinkedIn has none. The real difference is power. Google's server only reads your account, while Meta's can create, edit and delete campaigns.
- Google Ads MCP: official and read-only. Setup takes a developer token, a Cloud project and OAuth.
- Meta ads MCP: open to any Meta app since July 16, 2026, and it can delete campaigns and audiences.
- LinkedIn: no official server. Every LinkedIn Ads MCP is third-party.
- Bytown MCP: standard on Pro, advanced on Max, across Google, Meta, LinkedIn and ChatGPT Ads. Every change waits for your confirmation.
- Paid assistant plan: usually needed. Most assistants keep custom connectors behind a paid tier.
Published by Bytown: what each ads MCP server can do to your ad account in 2026, platform by platform. Google's official server only reads; Meta's can create, edit and delete campaigns; LinkedIn has none.
The question that matters is not which assistant you use. It is what you are letting that assistant do to an account with your credit card behind it.
MCP (Model Context Protocol) is the open standard that lets an AI assistant like Claude or ChatGPT use outside tools. An ads MCP server hands the assistant a set of tools for your ad account: some read reports, some change budgets, some delete things. Which tools a server exposes decides how much trust you are extending.
| Server | Built by | Reads your account | Changes things | Deletes things | Who can connect it | Status |
|---|---|---|---|---|---|---|
| Google Ads MCP | Google (official) | Yes | No: "Read-only (current release)" | No | Anyone who can supply a developer token, a Google Cloud project and OAuth credentials, and run the server | Docs updated Aug 3, 2026 |
| Meta ads MCP | Meta (official) | Yes, with breakdowns | Yes: campaigns, ad sets, ads, audiences, catalogs, A/B tests | Yes: campaigns, ad sets, ads, custom audiences | Any developer with a Meta app | Opened to all apps Jul 16, 2026 |
| Microsoft Advertising MCP | Microsoft (official) | Yes | No: pilot "focuses on read-only access" | No | Advertisers in the open pilot | Open pilot, June 2026 |
| Amazon Ads MCP | Amazon (official) | Yes | Yes: includes building Sponsored Products campaigns | Not stated | Open beta | Open beta from Feb 2, 2026 |
| LinkedIn Ads | No official MCP server. "LinkedIn ships no official MCP server." (Scalekit, Jul 29, 2026) | |||||
| Ryze AI MCP | Ryze AI (vendor, hosted) | Yes: 8 Google Ads read tools | Yes: 6 approval-gated write tools | Yes, approval-gated (runRawMutateDelete) | Anyone, sign in with Google; no developer token | Page updated Sep 3, 2026 |
| Bytown MCP | Bytown (vendor, hosted) | Yes: Google, Meta, LinkedIn and ChatGPT Ads accounts you connect | Yes, staged: you see the before and after, then confirm | Some, confirmed first: removes Google ads and ad extensions, archives Meta campaigns | Standard MCP on Pro, advanced MCP on Max; sign in with your Bytown account | Live in beta |
Sources for every row are listed at the end of the page. Vendor rows describe each vendor's own published page.
Google's official server only reads
Google's server is honest about its limits. Its developer page lists the mode as "Read-only (current release)" and exposes three tools: list_accessible_customers, which lists the accounts you can reach; search, which runs Google Ads Query Language (GAQL) requests "to fetch resource metrics, budgets, and status"; and get_resource_metadata, which describes the account's structure.
Here's the uncomfortable part for a small business: it is a developer tool. The setup asks for a "Developer token: Your unique 22-character access string", a Google Cloud project ID and OAuth credentials, and the server runs on your own machine or on hosting such as Google Cloud Run.
If you have never heard of a developer token, this server is not built for you yet.
What it is good for: asking questions of your data. "How did my campaigns do last week" works. "Pause the campaign that wasted money" does not, by design.
Meta's official server can change and delete
Meta went much further. According to PPC Land's report of Meta's July 16, 2026 announcement, agents "can create, edit, and delete campaigns, ad sets, and ads", and "Custom Audiences can be created, updated, and deleted through the same interface." Catalog work, performance reporting with breakdowns by age, gender and platform, signal diagnostics and A/B tests sit on the same server.
Here's where it gets interesting:
The change on July 16 was who gets access: Meta opened the server "to any developer holding a Meta app". That makes it infrastructure other tools will build on, which is useful, and it means a lot of new software will soon be able to edit and delete Meta campaigns on someone's behalf.
So what does that mean for you? Before you connect any tool built on Meta's server, ask one question: what happens between the AI deciding to change something and the change going live?
Meta's announcement, as reported, describes the capabilities. It does not describe an approval step for you to rely on, so that step is up to the tool you connect through.
Microsoft and Amazon
Microsoft's server is in an open pilot. Its product page promises "real-time performance insights and full campaign audits" from Copilot, Claude, ChatGPT and other assistants, and trade coverage from June 19, 2026 says the pilot "focuses on read-only access."
Amazon's server entered open beta on February 2, 2026. Amazon's announcement goes beyond reading: it describes a tool that expands a campaign "to another country with a single prompt" and another "to create end-to-end Sponsored Products campaigns in a single workflow." Its announcement does not say whether the server deletes.
LinkedIn has none
LinkedIn is the gap. Its developer catalog lists REST APIs, not an MCP server.
Scalekit's July 2026 review puts it flatly: "LinkedIn ships no official MCP server. Every 'LinkedIn MCP' available in July 2026 is community-built or vendor-hosted."
In other words:
If a tool offers you LinkedIn Ads inside Claude or ChatGPT, it is a third party wrapping LinkedIn's API.
Hosted vendor servers: Ryze AI and Bytown
Vendor-hosted servers exist because the official ones are either developer tools (Google), read-only (Microsoft), or full power with the approval step left to whoever builds on it (Meta). A hosted server takes care of the setup and decides what the AI may do.
Ryze AI runs a hosted server that, per its own page, connects to Google Ads "with no Google Ads API developer token, no Google Cloud project and nothing to install." It lists 14 Google Ads tools, 8 that read and 6 "approval-gated" write tools, including runRawMutate and runRawMutateDelete, and says it offers "150+ tools in total across Google Ads, Meta Ads, TikTok, GA4, Search Console and Shopify on the same endpoint."
Bytown's MCP is live in beta at ads.bytown.co/mcp and works with Claude, Claude Code, ChatGPT, Codex, Cursor, Antigravity and the Gemini CLI. Reads run instantly. Any change, whether a budget, a keyword or a piece of creative, "is staged with an exact before-and-after, and runs only after you confirm it", and pending approvals also appear in your Bytown dashboard.
It exposes the same tools as the assistant inside the Bytown app: 214 of them in September 2026, across Google, Meta, LinkedIn and ChatGPT Ads, and it comes with every Bytown plan: standard MCP on Pro, advanced MCP on Max.
That includes removals and money settings: it can remove a Google ad or ad extension, archive a Meta campaign and set a Meta account spending cap, each staged for your confirmation like any other change. It can also read Google Ads billing details.
It sees only the ad accounts you have connected to Bytown, "with the same permissions as your own login."
Full disclosure:
Bytown is not neutral here. Its MCP is part of the product, and this page is on Bytown's site. The table above states what each vendor publishes about itself, so you can check every row.
Which one should you connect?
It depends on what you want the assistant to do, and on how much of the setup you want to own.
| If you want to | Look at | Watch out for |
|---|---|---|
| Ask questions of your Google Ads data and you are comfortable with developer setup | Google's official server | Developer token, Cloud project and OAuth; it cannot change anything |
| Build your own tool that manages Meta campaigns | Meta's official server | It can delete; the approval step is yours to build |
| Analyse Microsoft Advertising data from Copilot or Claude | Microsoft's pilot server | Read-only during the pilot |
| Run Google and Meta from Claude or ChatGPT with no developer setup | A hosted vendor server (Bytown, Ryze AI) | Read what it may change and delete, and how approval works, before connecting |
What's actually important?
The honest answer for most small businesses: you do not need an MCP server to run ads well. You need one if you already work in Claude or ChatGPT all day and want the account inside that window. If you do connect one, pick the one whose limits you understand.
Three questions before you connect any ads MCP server.
- Can it change things, or only read?
- If it changes things, does anything run without your confirmation?
- Can it delete, and can it touch billing?
If the answer to any of them is "I don't know", do not connect it to the account that pays your bills.
Common questions
Is there an official Google Ads MCP server?
Yes. Google publishes an official Google Ads MCP server. Its developer page lists its mode as "Read-only (current release)", with tools to list accounts, run GAQL queries and read resource metadata. It needs a developer token, a Google Cloud project and OAuth credentials.
Can an AI assistant delete my Meta ad campaigns?
Through Meta's official ads MCP server, yes: PPC Land's report of Meta's July 16, 2026 announcement says agents "can create, edit, and delete campaigns, ad sets, and ads." Whether that happens without your approval depends on the tool you connect through it.
Does LinkedIn have an MCP server for ads?
No official one. As of Scalekit's July 2026 review, LinkedIn's developer catalog lists REST APIs only; every LinkedIn Ads MCP is built by a third party.
Can the Bytown MCP change my campaigns without asking?
No. Per Bytown's MCP page, every change is staged with an exact before and after and runs only after you confirm it. That applies to removals and money settings too, such as removing a Google ad, archiving a Meta campaign or changing a Meta spending cap.
Do I need a paid Claude or ChatGPT plan to use an ads MCP server?
Usually. Most assistants put custom connectors behind a paid plan, and ChatGPT adds them through developer mode. That cost is on the assistant's side, separate from the ad platform or the MCP vendor.
Sources
- Google Ads API docs, Google Ads MCP server (last updated Aug 3, 2026): developers.google.com
- PPC Land, "Meta opens ads MCP to any app, cutting integration code to zero" (on Meta's Jul 16, 2026 announcement): ppc.land
- Microsoft Advertising, MCP server product page: about.ads.microsoft.com; PPC News Feed, "Microsoft Advertising MCP Server in Open Pilot", Jun 19, 2026: ppcnewsfeed.com
- Amazon Ads, "Amazon Ads MCP Server" open beta, Feb 2, 2026: advertising.amazon.com
- Scalekit, "LinkedIn MCP vs API for AI Agents (2026)", Jul 29, 2026: scalekit.com
- Ryze AI, Google Ads MCP page (updated Sep 3, 2026): get-ryze.ai/google-ads-mcp
- Bytown MCP: bytown.co/mcp
Run your ads from Claude, with a confirm button
Bytown's MCP reads and changes your Google, Meta, LinkedIn and ChatGPT Ads accounts and stages every change with a before and after. Nothing runs until you confirm.
Connect Bytown to Claude Standard MCP on Pro, advanced MCP on Max. Compare plans.