Newsroom live
AFFHOUSE
Get the brief
Latest
Sweepstakes casino bans: what operators and affiliates facedStake sponsorship case: Everton, UFC, F1 and KickSportsbook marketing: what DraftKings and FanDuel spendSaaS Affiliate Programs: A Blogger's $680K Semrush CaseRobux gambling ruling: why affiliates are exposedPrizePicks: how DFS pick'em became a $2.5bn US businessPopunder Ads Case: $30K in MENA, 1,536 iGaming FTDsPlayer lifetime value case: $9,125 in CPA, $61,376 GGR
Knowledge base9 min read

Meta Ads MCP: what an AI agent can do in your ad account

Meta Ads MCP lets Claude, ChatGPT and other AI clients read and edit a Meta ad account. Here is what it covers, how to connect it, and what to keep manual.

Published Updated

Hand passing a brass key with engraved brand tag to a robotic gripper

Meta Ads MCP is Meta's hosted server that lets an AI client such as Claude or ChatGPT read and manage a Meta ad account through permission-scoped tools instead of Ads Manager screens. Third-party servers do a similar job. For a media buyer, the real question is which tasks to hand to the agent and which to keep.

Updated October 29, 2026. Meta announced the connectors in open beta on April 29, 2026, and said on July 16 that tool access is being enabled gradually, so tools and screens may differ.

What is Meta Ads MCP

Meta Ads MCP connects an AI assistant to your Meta ad account through the Model Context Protocol, an open standard for plugging AI applications into outside tools. A server publishes a list of actions, the AI client picks the ones that fit your plain-language request, and you see the result in the chat.

Meta's server lives at https://mcp.facebook.com/ads. It is one of two Meta ads AI connectors:

  • Ads MCP server. Hosted by Meta. Any MCP-compatible AI client can attach to it and turn a typed sentence into tool calls against the account.

  • Ads CLI. A command-line interface for Meta's Marketing API, run from a terminal or a script. Meta aims it at developers, automated jobs and coding agents.

The MCP server works only with the permissions you grant. The practical change is who types the request: you, in plain sentences, not a developer.

Official Meta server vs third-party servers

Open-source and commercial servers wrapped the Marketing API before Meta shipped its own. What differs is who holds your access.

Meta ads MCP serverMeta Ads CLIThird-party MCP server
Who runs itMetaYou, on your machine or serverA vendor, or you if self-hosted
How you sign inFacebook Login for Business (OAuth)System user access tokenVendor account, or your own Meta app and token
Where access livesHeld by your AI client after you approve it with Facebook LoginWherever you store the tokenWith the vendor, or in your local setup
What you trustMeta and the AI clientYour own scriptsMeta, the AI client and the vendor's code and security

The best-known third-party option is Pipeboard's open-source server, which shows up in searches for "pipeboard meta ads mcp" and "meta ads mcp github". Its README calls it an independent project on Meta's public APIs and says Pipeboard is a badged Meta Business Partner. It offers hosted and local versions, asks for explicit confirmation on every write, and uses the Business Source License 1.1. Tool lists differ between servers, and any third party means one more holder of a token to an account that spends real money.

What an agent can do in Ads Manager

Meta groups the server's tools into seven areas: reporting, ad creation and management, catalogs, signals and datasets, help and troubleshooting, A/B tests and lift studies, and activity logs. Sorted by job:

Reading and diagnosing (safe default)

  • Change history. Activity logs show who changed what and when, the first place to look after a cost jump.

  • Competitor research. A search tool queries the public Meta Ad Library for live ads by brand or keyword.

  • Performance data. Spend, impressions, CTR and ROAS for any date range, split by age, gender and platform. From that data the agent can compare periods and flag ad sets whose numbers moved.

  • Help lookups. The agent can cite Meta Business Help Center articles when it explains a delivery problem.

  • Signal quality. Dataset tools report signal health and quality, a quick check on whether your pixel and Conversions API setup send Meta clean events.

  • Catalogs. For product ads, the agent can trace feed issues.

Drafting and editing

The same connection writes. It can create campaigns, ad sets, ads and creatives (mainly single-image link ads), edit existing objects, build and update custom audiences, and set up A/B tests and lift studies. Meta's documentation says write tools create these objects paused. A separate activation tool switches them on, Meta warns that activation starts spending budget, and the AI client asks for confirmation first.

Paused by default covers new objects only. Edits to live campaigns, deletions and audience changes need their own controls.

What should stay manual

A simple ladder keeps the agent useful without handing it the wallet:

  1. Read: reports, diagnostics, audits. Let the agent run these freely.

  2. Draft: new campaigns, ad sets and ads in paused status. Let the agent build them, then review in Ads Manager.

  3. Change: budget, bid and targeting edits on live spend. Allow only with a named human approving each one.

  4. Never: activation without review, deletions, and anything touching billing, user roles or policy status.

The reason is scale: a wrong manual edit hits one ad set, while an agent working through a list repeats it on every row. It also cannot see your commercial context, such as an offer cap or a creative legal has not cleared.

Two items need care. Custom audience updates upload hashed customer data, which raises consent questions an agent cannot answer. And gambling advertisers stay inside Meta's own rules. Meta allows online gambling and gaming ads only after the ad account has received Meta's authorization, the rule also applies to affiliate and aggregator landing pages, and those ads cannot target people under 18. A connector does not change your policy status. This is a summary of Meta's published policy, not legal advice.

For a media buyer, the agent replaces report-pulling and first drafts, not judgement.

How to connect Meta Ads MCP to Claude

Meta's Get started page documents Claude Code and ChatGPT, both with a Meta app ID as the OAuth client ID. It also says owning an app is not a prerequisite, so the Claude app may look different. The flow below is the general pattern for Claude's custom connectors, not a Meta-published walkthrough.

  1. Point Claude at the server. Add a custom connector with the Meta Ads MCP server URL, https://mcp.facebook.com/ads (confirm it on Meta's Get started page).

  2. Approve with Facebook Login for Business. Use the Facebook account that already manages the ads, and grant access only to the businesses the agent should see.

  3. Check the permissions. Meta's Get started page lists the minimum as ads_mcp_management plus ads_read or ads_management. Catalog, business, Pages and Instagram permissions are optional.

  4. Run a read first. Meta's suggested opening request is "List my ad accounts."

In Claude Code, one claude mcp add command carries your Meta app ID; ChatGPT takes the server URL with OAuth and the same app ID. The app's redirect URL must match the client. Agencies that connect other businesses' accounts through their own app need Meta's App Review for the MCP permission.

To revoke access, disconnect the connector in your AI client. If it still appears among the apps and integrations in your Facebook or Business settings, remove it there too. Meta's MCP pages do not describe this step, so check the current menu names.

Tokens, permissions and security

An OAuth scope is the list of things you allow an app to do on your behalf, and it is the main safety control you have. Grant the narrowest set the job needs.

  • Start read-only. For reports, grant ads_read, not ads_management. Add write access later, for one account.

  • Limit the login. Use a Facebook user that reaches only the ad accounts the agent should touch.

  • Use portfolio rules. Since July 16, 2026, anyone with full control of a business portfolio can set ads MCP server rules that govern what AI agents may do, from budget changes to catalog updates.

  • Keep tokens out of chats. The OAuth route needs no manual token. If you do use a system user token, Meta accepts only employee-role ones; admin-role system user tokens cannot reach the server. Never paste a long-lived token into a chat or shared config; treat Ads CLI and self-hosted server tokens like passwords.

  • Watch for prompt injection. The agent reads text it did not write, such as Ad Library results. Hidden instructions there can steer a model, so writes need a human.

  • Keep confirmations on. Ads CLI flags that suppress prompts do not belong on a live account.

Limits and gaps

The connector launched in beta on a rate-limited API. The Marketing API throttles calls per ad account and per app and limits rapid edits to one object, so an agent looping through changes will hit errors. Coverage is Meta only, Meta says tool access is rolling out gradually so accounts differ, and the connector does not change your account's policy status or ad authorizations.

Example workflow: a weekly account review

A Monday routine that keeps the agent on the read and draft rungs:

  1. Check what changed. "Who edited budgets, bids or targeting in the last 7 days, and on which campaigns?"

  2. Catch overspend. "Which campaigns spent more than 20% over their daily budget yesterday?"

  3. Rank ad sets. "Compare the last 7 days with the 7 before for every active ad set: spend, CTR, cost per result, frequency. Biggest movers first."

  4. Check signals. "How healthy is the signal quality on the dataset this account optimizes for?"

  5. Draft tests. "In the ad set with the lowest cost per result, create three new ads with different hooks and leave them paused." Ask for angles that differ in hook and format, because creative variety drives delivery under Meta Andromeda.

  6. Human step. The buyer checks the drafts in Ads Manager against offer terms and policy, and activates what passes.

Treat the agent as a fast analyst whose edits you review.

FAQ

What is MCP in Facebook ads?

MCP in Facebook ads means connecting an AI assistant to a Meta ad account through the Model Context Protocol. The assistant calls tools on an MCP server, such as Meta's at mcp.facebook.com/ads, to pull reports, run diagnostics and draft campaigns without clicking through Ads Manager.

Is there an MCP for Facebook, and does Meta have one for Claude?

Meta has an official MCP for Facebook and Instagram ads. The Meta Ads MCP server was announced in open beta on April 29, 2026, alongside Ads CLI. It works with Claude and other MCP-compatible AI clients. Third-party servers, such as Pipeboard's open-source project, also exist.

How do I add Meta Ads MCP to Claude?

To add Meta Ads MCP to Claude, create a custom connector with the server URL from Meta's documentation and approve it through Facebook Login for Business. Grant only the businesses and permissions you need, then test with a read-only request.

Is it safe to give an AI agent access to my ad account?

Giving an AI agent access to your ad account is reasonably safe for reading data and risky for live changes. Grant read-only permissions first, limit the login to the accounts you need, set portfolio rules, and keep a human approval before activation.

Can Meta Ads MCP launch or change campaigns on its own?

Meta Ads MCP can create and edit campaigns, but new campaigns, ad sets and ads start paused. Switching them on takes a separate activation tool, and Meta says the AI client asks for confirmation first. Edits to live objects are not paused, so restrict them with permissions and portfolio rules.

Does Meta Ads MCP work with ChatGPT or only Claude?

Meta Ads MCP works with ChatGPT as well as Claude; Meta describes it as compatible with any MCP-compatible agent. Do not confuse it with ads inside ChatGPT: managing Meta campaigns from ChatGPT and buying ad space in ChatGPT are separate products.

Sources

Platform documentation checked October 29, 2026; gambling policy checked October 28, 2026.

Keep reading

All knowledge base →