A year ago, “connect Claude to your ad account” meant a weekend hack and a sketchy GitHub repo.
Now Meta, TikTok, Google, Microsoft, Amazon and Snapchat all have an official one.
Some of them read your reports. Some of them can launch campaigns and spend your money while you sleep. That’s a fairly important difference, so let’s sort it out.
Last checked on 2 November 2026. This moves every week, and I’ll keep the post updated.
What an MCP server is, in plain words
MCP stands for Model Context Protocol. Anthropic introduced it in November 2024 as an open standard. OpenAI adopted it a few months later, and today ChatGPT, Gemini and Microsoft Copilot speak it too.
Think of it as USB-C for AI. The ad platform builds one plug, and any AI assistant that speaks MCP can use it. You connect, log in with your normal ad account, and ask things like “which ad sets wasted money last week?” in plain English.
The plug decides what the AI is allowed to do. Some plugs are read-only. Some have write access. That’s the whole game.
The quick answer
| Platform | Official MCP server? | What your AI can do | Status |
|---|---|---|---|
| Meta (Facebook, Instagram) | Yes | Read and write: campaigns, ad sets, ads, audiences, A/B tests | Open beta since April 2026 |
| TikTok | Yes | Read and write: campaigns, budgets, targeting, creatives | Announced May 2026 |
| Amazon | Yes | Read and write, including billing data | Open beta since February 2026, partners with API access only |
| Google Ads | Yes, open source | Read only | Experimental since October 2025 |
| Microsoft Advertising | Yes | Read only | Open pilot since June 2026 |
| Snapchat | Yes | Read only, write access “planned” | Live since August 2026 |
| Not yet | Third-party tools only |
Two camps. Meta, TikTok and Amazon let the AI touch your campaigns. Google, Microsoft and Snapchat let it look, not touch.
Meta: the one that can actually spend money
Meta opened its Ads MCP server in April 2026, starting with Claude and ChatGPT. In July it opened it to any developer with a Meta app. Agencies working on other businesses’ accounts have to pass an app review first.
It can create and edit campaigns, ad sets and ads, build custom audiences, run A/B tests and pull reports. Basically everything you do in Ads Manager, minus the clicking.
What Meta’s docs don’t describe: spend caps, “are you sure?” steps or new ads paused by default. The safety net is whatever your AI app does before it acts. Sleep well.
Also worth knowing: Digiday reported people online claiming their accounts got restricted after letting AI run their Meta campaigns. Nobody has confirmed a link. I’d still start read-only.
There are popular third-party Meta servers too, like Pipeboard. Fine tools, but you’re handing your ad account access to one more company. Read the fine print.
TikTok: write access, plus an app store for agents
TikTok announced its MCP server in May 2026 and followed up with the Agentic Hub in June, a marketplace of AI “skills” from partners like HubSpot and Wix. It’s in Claude’s connector directory as well.
The write access is real: creating campaigns, pausing ad groups, changing budgets and targeting, creatives and reporting. TikTok’s own help page talks about agents working “without human intervention at every step”. Exciting or terrifying, depending on your budget.
Google Ads: read-only and open source
Google’s MCP server came out in October 2025 as an experimental, open-source project. Google is very clear about it: “This implementation is strictly read-only.” It runs reporting queries. No bid changes, no pausing, no new ads.
It’s also more hands-on. You usually host it yourself and need Google Ads API access through a Google Cloud project. Fine for an agency with a developer, a lot for a solo buyer.
Microsoft and Amazon: opposite ends
Microsoft moved its MCP server to an open pilot in June 2026. Read-only, works with Copilot, Claude and ChatGPT. Great for audits, useless for launching.
Amazon went the other way: read and write, including tools that build a whole Sponsored Products campaign from one prompt and then wait for your approval. The catch: it’s only open to partners with Amazon Ads API access.
Snapchat: read-only, but the best guardrails
Snap announced its MCP server in June and switched it on in August 2026. Right now an agent “can read your Snapchat Ads data but cannot create, update, or delete anything”. Write access is “planned for a future release”.
What Snap did better than everyone else is control. An admin approves each AI tool for the whole organisation, every team member logs in separately with their own permissions, and access can be revoked at any time. Snap even warns about prompt injection, which is a fancy way of saying “a webpage can trick your AI into doing something dumb”.
The read-only part matters for anyone running volume on Snapchat. You can’t bulk-upload creatives or launch campaigns through the official MCP yet. That still goes through Snap’s Marketing API, which is what Adkorly runs on.
LinkedIn: not yet
As of early November 2026, LinkedIn hasn’t announced an official Ads MCP server. There are third-party ones, but LinkedIn’s API access is tight. Some tools work by driving your logged-in browser, which breaks LinkedIn’s user agreement. Skip those.
How I’d use this today
- Start read-only. Connect reporting first. Ask the questions you’d ask a junior buyer on Monday morning: what spent, what converted, what died.
- Give the AI its own login. Its own user, with the smallest permissions that work. If something goes wrong, you revoke one login, not your whole account.
- Let it draft, not publish. On Meta and TikTok, have the agent create things paused and check them yourself.
- Cap the damage. Set spend limits on the ad account itself. An agent can’t blow a budget that isn’t there.
- Keep a log. Know what the AI changed and when. Meta has an activity log for exactly this. Use it.
Treat your AI like a very fast new hire. Smart, eager, and absolutely not getting the company card on day one.
Where this is going
Every platform is building its own plug. Great, until you run Snapchat, TikTok and Meta at the same time and your agent needs three logins, three billing setups and three sets of rules.
That’s the layer we’re building at Adkorly: one place where your agent gets ad accounts, funds them from one wallet, and works within spend limits you set, across platforms. More on that in the Korly post.
Which ads MCP server have you actually connected, and did it do anything useful? I’m collecting real stories, good and bad. Tell me.
Cheers,
Johannes
