Guide

Meta Ads MCP: connect Facebook and Instagram ads to Claude or ChatGPT

Meta's official ads MCP server (no fee during the beta), Windsor.ai, Pipeboard or a read-only data feed: four ways to manage Meta ads from Claude or ChatGPT, compared on setup, price, write access, safety rules and data.

Updated 2 Nov 2026

Meta now runs its own MCP server for ads. Connect it to an AI chat such as Claude, and you can ask about your Facebook and Instagram campaigns in plain language, or have the AI build a campaign for you to review. In Claude this takes a few clicks; in ChatGPT the current route runs through developer mode and needs more setup. This guide explains what Meta’s official server can do, how to connect it safely, and when a third-party connector such as Windsor.ai, Pipeboard or a read-only data feed fits better. Facts come from Meta’s Business Help Center and developer documentation and from each vendor’s own pages as of 2 November 2026, linked under Sources at the end. We read the documentation; we did not run the connectors on a live ad account.

What is the Meta Ads MCP server?

MCP, short for Model Context Protocol, is an open protocol that lets a compatible AI client call outside tools and fetch data while it answers you. Meta’s ads MCP server is that tool for your ad account. Many people search for it as “Facebook Ads MCP”; it covers Facebook and Instagram ads alike. It is hosted by Meta at https://mcp.facebook.com/ads, has been in open beta since 29 April 2026, and, according to Meta’s developer documentation, “works with any MCP-compatible AI agent”. Meta calls the family “Meta ads AI connectors”; the other member is a command-line tool for developers and coding agents.

The server groups its tools into seven areas: reporting, ad creation and management, catalogs, signals and datasets, help-center search, A/B tests and conversion lift studies, and activity logs. It can also search the public Ad Library, manage custom audiences and boost Instagram posts. Meta notes that “you may not have access to all of these tools and features yet”.

Which option is right for you?

Only Meta ads, and you want the official route: Meta’s ads MCP server. Meta lists no fee during the beta, Claude connects with a Facebook login and no Meta developer app, and the server reads and writes.
Meta ads next to Google Ads, GA4 or other platforms in one chat: Windsor.ai. Free for one data source; $19 a month billed yearly ($23 monthly) for three. It can also change campaigns if you let it.
Several Meta ad accounts with read-only tokens for reporting: Pipeboard. Free for two ad accounts and 30 tool runs a week; Pro from $29.90 a month.
Numbers only, no write access at all: a read-only data feed such as Supermetrics (from $44 a month billed yearly, with the AI chat as its destination) or Coupler.io (its AI connection needs at least the Professional plan).

How do the four options compare?

Checked	Meta ads MCP server	Windsor.ai	Pipeboard	Read-only feeds (Supermetrics, Coupler.io)
What it is	Meta’s own hosted MCP server, open beta	Hosted MCP connector for 350+ data sources	Hosted MCP server for Meta ads and other ad platforms	Reporting tools with an AI-chat connection
Setup	Add a custom connector with Meta’s URL, log in with Facebook	Add its connector from Claude’s directory or the ChatGPT app, sign in	Add Pipeboard’s URL as a connector, log in, connect Facebook	Sign up, pick Facebook Ads as source and the AI chat as destination
Platforms	Meta only	Meta, Google Ads, GA4, TikTok, LinkedIn and more	Meta and other ad platforms	Many sources
Can change ads	Yes: campaigns, ad sets, ads, budgets, targeting, audiences, catalogs	Yes, per its documentation: create, pause, enable, update, budgets, boost posts	Yes, per its documentation: create (paused), status, bids, targeting	No write actions documented
Safety controls	Rules per ad account in Meta Business Suite, plus your AI app’s tool approvals	Your AI app’s tool approvals	Your AI app’s tool approvals; read-only tokens on paid plans	Read only by design
Price	No fee listed during the beta	Free for 1 source; from $19 a month billed yearly	Free for 2 ad accounts; Pro $29.90 a month	Supermetrics from $44 a month billed yearly; Coupler.io AI access from the Professional plan
Tested by bto.ai	Documentation only	Documentation only	Documentation only	Documentation only

How to connect Meta Ads MCP to Claude

The labels below were documented on 2 November 2026 and may change.

Add the connector. In Claude, open Customize, then Connectors, click + and Add custom connector. Enter https://mcp.facebook.com/ads as the URL and click Add.
Log in with Facebook. Click Connect and approve the request in Facebook Login. When it asks which business portfolios, Pages and Instagram accounts to share, choose specific current ones instead of “all current and future”.
Switch it on in a chat. In a new chat, open +, then Connectors, and make sure the Meta connector is on.
Set the tool permissions. In the connector’s settings, allow the reading tools and set every tool that creates, edits, activates, uploads or deletes to ask for your approval. Never pick “always allow” for those. See the safety section below.
Start with a read-only question, such as “Summarize the top-spending campaigns in my ad account over the last 30 days.”

On Team and Enterprise plans, an organization owner may have to add the custom connector first. Check Anthropic’s current help pages for the connector limits of your plan.

How do you connect Meta Ads MCP to ChatGPT?

In ChatGPT, custom MCP servers run through developer mode. Meta’s documentation describes that route with your own Meta developer app and its app ID for the OAuth login, which is a developer task, not a few clicks. If you are not comfortable creating a Meta app, use Claude for Meta’s server, or a connector that is available as a ready-made ChatGPT app, such as Windsor.ai. Check OpenAI’s current instructions for your plan before you start.

What can you ask once Meta ads are connected?

Wasted spend: “Which ad sets spent more than $100 in the last 14 days with a cost per result at least 50% above the account average? Only analyze, change nothing.”
Creative fatigue: “Show ads whose CTR fell by more than 30% while frequency rose above 3 over the last four weeks.”
What changed: “List every change to budgets, targeting or status in this ad account over the last seven days, and who made it.”
Competitors: “Search the Ad Library for active ads from <competitor> in the US and summarize their offers and formats.”
Across platforms (Windsor.ai only): “Compare cost per conversion from Meta Ads and Google Ads last quarter. List each platform’s conversion definition and attribution setting first.”

Treat every answer as a draft. Before you change spend because of it, check the ad account ID, currency, time zone, date range, attribution setting, metric definition and totals in Ads Manager. Meta and Google count conversions differently, so a cross-platform comparison is only as good as the definitions behind it.

Can Claude or ChatGPT change your Meta ads?

Yes, with Meta’s server and with Windsor.ai or Pipeboard. Meta’s server can create and edit campaigns, ad sets and ads, change budgets and targeting, create and delete custom audiences, upload hashed customer lists, and manage catalogs. New campaigns, ad sets and ads are created paused, but a separate tool, ads_activate_entity, switches them on and “starts spending budget”. Edits to a campaign or ad set that is already running take effect like an edit in Ads Manager, without any activation call.

Two layers of control decide what the AI can actually do:

Rules in Meta Business Suite. Anyone with full control of a business portfolio can open Settings, then Integrations, then Ads MCP server, pick an ad account or catalog and block actions such as creating campaigns, editing targeting or changing delivery status. For budgets, you can block all changes or only budgets above an amount you set. Meta enforces these rules on its server, whatever the AI app is told. Meta lists the rules as limited availability, so the page may not appear for your business yet.
Tool approvals in your AI app. In Claude, each tool of a connector can be set to always allow, ask for approval or block. With approval on, Claude shows the call and its values before it runs.

Before you let an AI write to a live account:

If the rules page is available, block delivery-status changes and set the budget ceiling below the most you are willing to lose.
Set every create, edit, activate, upload, audience and delete tool to ask for approval.
Before you approve, check the ad account, the campaign or ad set name, its current status, the budget type (daily or lifetime) and the exact new amount. An approval screen with IDs only is a reason to stop and ask the AI for names.
Connect only the ad accounts you need.

Writing “analyze only” in a prompt helps, but it is an instruction, not a lock.

Can using an AI agent get your ad account banned?

Meta’s help center says it is “unlikely that an account would be banned simply for using it”, and that bans come from breaking its advertising standards, such as non-compliant ads. That is not a guarantee: ads and account activity created through an AI still have to follow Meta’s policies. We found no comparable statement in the third-party documentation we reviewed; Windsor.ai and Pipeboard connect through Meta’s login and Marketing API, which does not exempt anything from those policies. Also note that significant edits can send an ad set back into the learning phase, so batch changes deliberately instead of letting an AI make many small ones.

When is Windsor.ai the better choice?

Meta’s server only covers Meta. If you want one chat to answer questions across Meta Ads, Google Ads, GA4, TikTok or LinkedIn, a multi-platform connector saves you from running a separate MCP server for each. Windsor.ai (affiliate link) is listed in Claude’s connector directory and as a ChatGPT app, so you add it from there instead of pasting a URL. According to its documentation, its write actions for Meta include creating, pausing, enabling and updating campaigns, ad sets and ads, setting budgets and boosting posts, on every plan including the free one; we did not verify them on a live account. The same safety advice applies: set its execute_action tool to ask for approval.

The limits: Windsor.ai is one more company that processes your ad data, and it counts sources strictly. Facebook Ads (which includes Instagram ads), Facebook Pages and Instagram organic data are three separate data sources, so Meta Ads plus Google Ads already needs the Basic plan after the 30-day trial. Its terms rule out refunds, so use the trial to test your exact questions before you pay. Our Windsor.ai review covers plans, terms and user reviews, and the Google side is in our guide Google Ads MCP: connect Google Ads to Claude or ChatGPT.

Is Meta Ads MCP free, and what do the alternatives cost?

Meta’s ads MCP server: Meta lists no fee during the open beta and has not published pricing for afterwards.
Windsor.ai: free for one data source and one account after the trial. Basic is $19 a month billed yearly ($23 monthly) for three sources and 75 accounts; Standard is $99 billed yearly ($118 monthly) for seven.
Pipeboard: free for two ad accounts and 30 AI tool runs a week. Pro is $29.90 a month for three ad accounts and unlimited runs, Premium $99 for ten. Yearly billing saves about 17%.
Supermetrics: Starter from $44 a month billed yearly ($55 monthly), with the AI chat as its one destination. Read only.
Coupler.io: has a free plan and a Starter plan from $24 a month billed yearly, but its MCP page names at least the Professional plan for the AI connection. Read only.

The AI chat is extra. Meta’s server has nearly 100 tools, according to a count by Meta ads specialist Jon Loomer, and long data questions use a plan’s limits faster, so a paid Claude or ChatGPT plan is realistic for regular use.

Where does your data go?

With Meta’s server, Meta hosts the MCP server, and the AI provider you connected receives the tool results it needs to answer. Both may keep logs under their own terms, and how the AI provider retains or trains on that data depends on your plan and settings. With a third-party connector, one more company processes the data. Windsor.ai says it runs its servers in Germany, keeps data in memory for up to 24 hours and on disk until 20 days without access, with no backups of customer data; that does not cover every sub-processor, so read its DPA and sub-processor list. Check the same documents for any other connector before you connect client accounts.

Custom audiences built from customer lists are personal data, even when the upload is hashed. Only upload or change one if the advertiser has authorized it and you have a legal basis, never paste raw customer data into the chat, and check Meta’s Custom Audience terms and the points in AI tools and the GDPR first.

Which one should you pick?

For Meta only, try Meta’s own server first if you can set the safety rules and are fine with beta software. It lists no fee during the beta and covers the most Meta features.
Add Windsor.ai when you want Meta and Google Ads, GA4 or other platforms in the same conversation, without code.
Choose Pipeboard if you manage several Meta ad accounts and want read-only tokens for reporting.
Choose a read-only feed such as Supermetrics or Coupler.io if the AI should never be able to change anything, or if Meta’s rules page is not available to you yet.