What Meta's official Ads MCP does and does not do, six months in

Meta's free Ads MCP server, six months after launch: what it handles well, the six problems media buyers report, and where a bulk workflow hits the wall.

Jeremy Chaker, Published

Founder, Adsap11 min read

Meta's official Ads MCP is a Meta-hosted Model Context Protocol server at mcp.facebook.com/ads. It gives an AI assistant such as Claude or ChatGPT a set of tools to read and write your Meta ad accounts. Meta opened it in beta on April 29, 2026, and opened it to developers' own apps on July 16, 2026. Six months in, here is what it does, what it does not do, and where a media buyer hits the wall.

I run Meta accounts every week and I build Adsap, a tool that connects to the same API. Read this with that in mind. Every claim below has a source at the end of the post.

What is Meta's official Ads MCP?

Meta calls it the ads MCP server. It is hosted by Meta, so nothing runs on your machine. You add it to any MCP-compatible assistant as a custom connector, log in with your Meta account, and the assistant receives a list of tools. Meta groups the tools into seven categories. Four cover the account: reporting, ad creation and management, catalog creation and management, and signals and datasets. Three cover the rest: help and troubleshooting, A/B tests and conversion lift studies, and activity logs.

Meta does not publish a tool count. In a Claude session in September 2026 we counted 97 tools. An independent explainer from August counted 29, so the set grows between versions. Meta's July post lists the write access: create, edit and delete campaigns, ad sets, ads, custom audiences and catalogs. The same post says nothing about rate limits, paused status or account-integrity rules.

DateWhat changedSource
April 29, 2026Open beta of Meta Ads AI Connectors: the MCP server at mcp.facebook.com/ads and a developer CLIMeta newsroom
July 16, 2026Developers can connect their own apps. Managing other businesses' accounts needs the ads_mcp_management permission at Advanced Access, through App ReviewMeta for Developers blog
July 2026A set of account-level rules appears in the toolsetOur Claude session
August 2026A per-account permission panel with tool-level switches, reported by one vendor blogadsuploader.com, single source
September 202697 tools counted in a Claude sessionOur Claude session

What does it do well?

Reporting is the strong part. Ask for spend, ROAS and cost per result by campaign for the last 14 days and you get a table in seconds. Breakdowns by age, placement or country work the same way. Single edits work too: pause an ad set, change a daily budget, rename a campaign.

Setup takes about 90 seconds and costs nothing, on every Claude plan including the free tier. Everything the connector creates lands paused, and the August explainer notes that no flag overrides that default. It can also read Meta's activity log for one account per request, so you can ask what changed yesterday and who changed it.

Example prompt
Show spend, results and cost per result by campaign for act_123 over the last 14 days, sorted by spend.

That is the sweet spot: one account, a question about the numbers, or one change you would otherwise click through in Ads Manager.

What did media buyers report over the summer?

Between June and August 2026 I collected the problems advertisers posted on Reddit, X and LinkedIn about the official connector. Six patterns came back. Each one below carries its source, and I flag the ones that rest on a single report.

1. It executes partially and reads incompletely

One buyer asked the connector for an ad with three primary texts and three headlines. On every run the first text option went missing, and it took a second call to add it. A freelancer managing 150 ad accounts tried to build a dashboard through Claude. He gave up because the connector "cannot download all the data every time". Neither case threw an error. That is the worst kind of bug in advertising: a report short of rows still looks like a report.

2. Tools are missing, so you finish in Ads Manager

A consultant hired to build a "type a prompt, get a campaign with creatives" workflow found two gaps. The connector had no tool for Advantage+ creative options and could not upload creatives into the account's library. His client's launch still ended in Ads Manager. An Italian agency answered the 150-account thread with what they did instead: a custom CLI, two months of work with Claude Code.

3. The advice is generic

The connector hands the model raw metrics. It knows nothing about margins, lead capacity or what you tested last quarter. An agency owner who tried it on several accounts wrote that the answers "are often inaccurate" without that context, and that it gave "bad strategic advice in general". Another reply put it more plainly: useful if you know the fundamentals, "otherwise extremely generic". Broad "analyze my account" prompts also burn through tokens. The standard advice in the June thread was to scope each request to one tight task.

4. Agencies still meet App Review

The launch message was no developer app and no review. That holds for your own ad accounts. Meta's July 16 post says that managing data on behalf of other businesses needs the ads_mcp_management permission at Advanced Access, which goes through App Review. The queue the connector removed comes back for the people who manage the most accounts. One caveat: that requirement is documented for the developer route. The docs do not say how an agency using the hosted connector with partner access is treated.

5. Advertisers are still afraid to connect it

In March 2026 a wave of accounts was banned after people pointed unofficial agents and raw API scripts at their ad accounts. The official server launched six weeks later, and the trust did not follow. "Has anyone used the official MCP without being banned?" is a thread title from late June. Replies ranged from "it's official, use it" to "test it on a throwaway account first". To be precise: I found no verified report of an account banned while using the official server in that window. The bans came from unofficial routes. Meta's automation restriction wave in July and August also hit accounts with no AI at all, so nobody can promise immunity either way.

6. Paused by default is a status, not an approval step

Paused by default is the connector's main guardrail. It is a seatbelt. What is missing is a review step, throttling, and a limit on how much an agent can stage in one go. The August explainer names the real risk: approval fatigue. An agent stages forty campaigns in a minute, a person reads three and switches on the batch. The author caps what he lets an agent stage per run at five. One vendor blog also reported that the August permission panel arrived with several write actions switched on per account. That is a single source, so treat it as a reason to open Business Settings, Integrations, Ads MCP Server and check the switches yourself.

Where does a media buyer hit the wall?

Put the six reports together and the wall has a shape. The connector is a thin bridge to the Marketing API, one request at a time.

  • No bulk. No spreadsheet view exists to build or edit fifty ads at once. Fifty ads is fifty requests, each one read back by you.
  • No preview before a write. You find out what the model sent by reading the result.
  • No per-action audit log across accounts. The activity-log tool covers one account per request.
  • No creative upload into the account's library, and no Advantage+ creative options, per the July report above.
  • Meta only. Your Google Ads accounts need another connector.

None of this matters at low volume. It matters the week you launch a product range across three markets.

What should you use it for?

Use it for what it is good at, and structure the request.

  1. Reporting on one account with a fixed date range and a fixed metric list.
  2. Diagnostics: pixel and dataset health, ad review status, the activity log for yesterday.
  3. Single changes you would otherwise click through: pause an ad set, change one budget, rename twenty campaigns to a convention.
  4. Reading the Ad Library, when the search tool answers. One August report says it errored on every query.

Keep write requests small enough to read in full. The connector will not do that for you.

Example prompt
List every ad set in act_123 with spend over $50 and zero results in the last 7 days. Do not change anything yet.

If you have not connected it yet, Meta's setup is a custom connector pointing at mcp.facebook.com/ads. The Adsap docs cover connecting the Adsap copilot in the same assistants, if you want the two side by side.

When is a bulk tool with guardrails worth it?

The official server is free and official, and for reporting it is good. The question is what happens the week you need to launch in bulk or run three accounts. Or the week you let an assistant write to a live account and want a review step in front of it. That is where a layer on top of the same API earns its price. Adsap is one such layer. It adds a spreadsheet-style editor for hundreds of ads, a preview before every write, and an audit log on every action. The same tools run in Claude, ChatGPT or Perplexity. The row-by-row comparison is on Adsap vs Meta's official Ads AI connector. The account setup is in the Meta Ads integration docs, and each guardrail is listed in the safety model.

Frequently asked questions

Yes. It is free on every Claude plan, including the free tier, and on any other MCP-compatible assistant. You pay for your assistant subscription, not for the connector.

No. It creates one entity per request and has no spreadsheet view. Bulk work means one request per ad, each one reviewed by you.

Meta's July 16, 2026 post says managing data for other businesses needs the ads_mcp_management permission at Advanced Access, which goes through App Review. That applies to the developer route. The docs do not say how partner access through the hosted connector is treated.

I found no verified public report of a ban tied to the official server between June and August 2026. The spring bans came from unofficial agents and raw API scripts, and the summer restriction wave also hit accounts with no automation at all.

Sources

Related posts

Bulk operations

How to bulk upload Facebook ads: three methods compared

Ads Manager duplication and its import spreadsheet, the Marketing API or a script, and a bulk tool with a spreadsheet editor, compared on setup, limits, creatives and review.

10 min read

Your next launch takes about 10 minutes

Launch Meta and Google Ads in bulk from a spreadsheet-style web app, or run the same tools from Claude, ChatGPT or Perplexity. Every change shows a preview first, and everything is created paused.

Start freeBook a Demo