Skip to content
AI stack, guide

Meta Ads MCP: a safe setup for AI agents

Last verified

Meta runs an official Ads MCP server that lets AI apps read reporting and manage campaigns through your own developer app. Connect it read-only first, give the agent the narrowest ad account access, require approval for every write, set account spend limits, and review the activity log weekly.

Meta Ads MCP guardrails
Meta Ads MCP guardrailsStates of an MCP connection to an ad account: read-only by default, a proposed write waits for a person, an approved write runs only within the spend cap, and every action is written to a log.STATES OF THE CONNECTIONRead-onlyreports, diagnosticsread · insightsAGENT PROPOSESWrite proposeddiff shown to youpropose · budgetPERSON APPROVESWrite within capstops at the spend capapprove · personwrite · in capREJECTEDDONEread · resultsACTION LOG: EVERY CALL, READ OR WRITEMeta Ads MCP guardrailsStates of an MCP connection to an ad account: read-only by default, a proposed write waits for a person, an approved write runs only within the spend cap, and every action is written to a log.STATES OF THE CONNECTIONRead-onlyreports, diagnosticsread · insightsAGENT PROPOSESWrite proposeddiff shown to youpropose · budgetPERSON APPROVESWrite within capstops at the spend capapprove · personwrite · in capread · resultsACTION LOG: EVERY CALLrejected, or done: back to read-only
Illustrative flow, simplified.
Start read-only. A write is a proposal until a person approves it, it runs only inside the spend cap, and every call, read or write, lands in the action log.

What the Meta Ads MCP server is

MCP (Model Context Protocol) is an open standard that connects AI apps to data and tools, such as ad platforms, BI and CRMs. The Meta Ads MCP server is Meta's hosted implementation for advertising. It runs at mcp.facebook.com/ads, and an AI client such as Claude or ChatGPT can call it to read data and act on an ad account.

Meta groups the tools into seven categories: reporting, ad creation and management, catalog, signals and datasets, help-center search, A/B tests and conversion lift, and activity logs.

Access runs through the app permission ads_mcp_management. Meta's announcement says that from 16 July 2026 any AI app can connect using your own developer app. Secondary sources disagree on how many tools the server exposes, so read Meta's tools page for the current list instead of trusting a number.

An MCP server acts with the permissions of the person who connects it. The agent can do whatever your login can do, so the safety work happens in the account and the client, before the first prompt.

Set up with your own developer app

Meta's documentation is the source for the exact screens and fields. This guide gives the order of work and leaves the click paths to Meta's Ads MCP Server overview and tools pages on developers.facebook.com.

  1. 1.Read the Meta pages first

    Open the Ads MCP Server overview and the tools page on developers.facebook.com. Note the current tool list and the permission each group needs. Do not rely on a tool count from a blog post.

  2. 2.Create your own developer app

    The server connects through a developer app that you own. Create it under the business that owns the ad account, so access stays under your control and can be revoked in one place.

  3. 3.Request the ads_mcp_management permission

    This is the permission Meta names for the server. Grant it only to the person or system user who will run the agent, and use an ad account role with the least access that still allows reporting.

  4. 4.Add the server to your AI client

    Register the hosted address as a remote MCP server in Claude, ChatGPT or another client that supports MCP. Turn on per-tool approval in the client if it offers it.

    Server address: mcp.facebook.com/ads
    Auth: your own developer app
    Client setting: ask before each tool call
  5. 5.Run a reporting-only session

    Start with a prompt that forbids changes and ask for a summary. Check the numbers against Ads Manager before trusting the agent with anything else.

    Read-only task. Use reporting tools only. Do not create, edit, pause or delete anything. Summarise spend, results and cost per result by campaign for the last 7 days.

Guardrails before the agent touches anything

Treat the first week as an audit of the agent, with no changes made to the account. The table lists the controls to set up and how to confirm each one is in place.

Guardrails for a Meta Ads MCP connection
GuardrailSettingHow to check
Read-only firstAllow reporting tasks only; no write tasks in the first weekAsk the agent to list the tools it used after each session
Narrow ad account accessGive the connecting user the lowest role that supports reporting, on one ad accountReview the roles on the ad account and remove anyone who does not need access
Approval for writesRequire a human click before any create, edit, pause or budget changeTry a harmless edit request and confirm the client stops and asks
Account spending limitSet an account spending limit that you can afford to lose in a bad weekConfirm the limit shows in the account settings
Separate test accountPractice write tasks on a low-budget or test account before a live oneKeep production and test credentials apart
Weekly log reviewRead the activity log each week, including changes the agent madeMatch each logged change to an approval you gave
Token hygieneKeep credentials out of prompts, chats and repositoriesRotate the credential if it ever appears in a transcript

Meta's tools page shows whether a reporting-only scope exists for the permission. If it does not, the role, the client approval step and the spending limit become your only brakes, so set all three.

Approval only works if the person approving reads what the agent proposes. A budget change on the wrong ad set looks the same as one on the right ad set unless you check the ids. Ask the agent to state the campaign, the old value and the new value before you accept.

Tasks that are safe to start with

Read tasks cannot change spend, so they carry the least risk while you learn how the agent behaves. They also show whether it reads your account structure correctly.

  • Weekly performance summaries by campaign, ad set and creative
  • Delivery diagnostics, such as which ad sets are limited by budget or stuck in learning
  • Comparing results across date ranges and flagging large swings
  • Checking signals and dataset health, such as whether purchase events arrive
  • Searching Meta's help center from inside the chat
  • Reading the activity log to list who changed what

A campaign export can also feed rules that need no account access at all. The ua-kill-scale-rules skill in the ua5er/user-acquisition-skills repository writes stop, scale and pacing rules from your CAC target and applies them to an export. Use rules like that to prepare changes, then make them yourself in Ads Manager.

Move to write tasks only after several clean read sessions. Begin with small, reversible changes, such as pausing one ad, and keep approval on.

Community MCP servers and their licences

Third-party servers exist for Meta Ads. They come from other publishers, not from Meta. A licence tells you what you may do with the code. It says nothing about whether the server is safe to hold your ad account token. The notes below come from a file list and pattern search checked on 15 November 2026. No scripts were executed, and write behaviour was not reviewed for any of them.

  • getmcpads-com/meta-ads-mcp-server: Apache-2.0. The scan found a doctor script that reads a token and a build step that uses child_process.
  • pipeboard-co/meta-ads-mcp: BUSL-1.1, which is source-available and not open source. It offers a hosted flow plus a local one, and it handles Meta tokens. Read the licence on its repository before use. This guide links nothing from it and copies nothing.
  • brandu-mos/konquest-meta-ads-mcp: MIT, open-core with a paid premium tier. The scan found 65 scripts and Meta tokens.
  • RoyAzran/mcp-ads: MIT. It covers Google Ads, Meta Ads and GA4 with a read-before-write skill. The scan found 77 scripts that reference SECRET in checks, which calls for a manual review.

If you use one, read the code that touches tokens, run it on a test account first, and confirm whether it holds your credentials locally or sends them to a hosted service. MIT and Apache-2.0 also require you to keep the copyright notice if you copy or redistribute the code.

Questions

Is the Meta Ads MCP server official?
Yes. Meta documents it on developers.facebook.com and announced it on facebook.com/business. The hosted address is mcp.facebook.com/ads. Community servers for Meta Ads are separate projects from other publishers, with their own licences and their own handling of your tokens.
Do you need your own developer app?
Meta's announcement says that from 16 July 2026 any AI app can connect with your own developer app. Access also needs the ads_mcp_management permission. Check Meta's overview page for the current requirements before you start, since they can change.
Can the agent spend money on its own?
It can if your role allows writes and the client does not ask first. Limit the risk with a low ad account role, approval for each write in the client, and an account spending limit. Start with reporting tasks only.
Should you pick a community server instead of Meta's?
It depends on the job. Start with Meta's server for reporting and diagnostics, because Meta documents its tools. Consider a community server only if you need a workflow Meta's tools lack, and then review its licence, its token handling and its write behaviour first.

Need creative testing and AI tooling tied to your app's cost per install? Start with a short audit.

See the service