Meta Ads MCP: how to connect Claude or ChatGPT to your Meta ad account
Meta's own ads MCP server, third-party connectors and open-source servers compared on who holds the credential and what the agent may write, with setup steps for Claude and ChatGPT.

A Meta Ads MCP server lets an AI client such as Claude or ChatGPT read and change a Meta ad account through the Model Context Protocol. As of November 2026 there are four routes: Meta's hosted server at mcp.facebook.com/ads with your Meta login, the same server through your own developer app, a hosted third-party connector, or a self-hosted open-source server.
You probably got here after a Monday like this one: the weekly Meta report took two hours again, someone pasted an Ads Manager export into ChatGPT, and the CFO asked why it takes a day to answer "what did we spend on Advantage+ last week". Connecting an AI client directly to the account fixes the two hours. It also hands a chat window the power to activate a campaign, so the setup deserves more care than a browser extension. Scalewell helps paid teams put a connection like this to work in their week, from the questions they ask it on Monday to the reporting it takes off their plate.
This guide is for the head of paid media or the PPC lead who will do the connecting. It surveys every public route we could verify in Meta's and the vendors' own documentation, gives a decision rule that favors none of them, and walks through the exact steps for Claude and ChatGPT. The route you pick decides who holds the credential, what the agent may write, and where the audit trail lives. The sibling guides cover the same question for Google Ads and LinkedIn Ads, because the access models differ enough to need separate articles.
Meta's own server removed the biggest barrier in April 2026
Meta launched its ads AI connectors, an ads MCP server plus an ads command line interface, in open beta on April 29, 2026, and the announcement says the MCP route needs "no developer credentials, API setup, or coding." The third-party servers in the table below, which many people still search for as a Facebook ads MCP, all need a Meta developer app or a Marketing API token, which is why the early guides read like onboarding documents for engineers.
The Model Context Protocol itself is the open standard Anthropic published on November 25, 2024, which the protocol site describes as "an open-source standard for connecting AI applications to external systems." Think of it as the plug shape: Meta's server is one appliance that fits it, and so is every third-party server below. Your AI client is the socket, and it decides whether a write action gets a confirmation prompt.
Meta's developer documentation lists the server's tools in seven groups, from reporting and ad management to catalogs and activity logs, and counting the tool pages on November 8, 2026 gives 89 tools. Windsor.ai's October 2026 test saw 98 tools from the live server against 89 on the documentation pages, so expect the count in your client to keep moving. The Business Help Center page still opens with "You may not have access to all of these tools and features yet," which is the open beta talking.
Four routes into a Meta ad account, compared
The four routes differ on credential, write scope and cost more than on what they can read. The table is what we could confirm on each provider's own pages on November 8, 2026, and where a page did not state something, the cell says so.
| Route | What it reads and writes | How it authenticates | Cost as documented |
|---|---|---|---|
| Meta hosted server, your login (docs) | Reporting, campaign creation and edits, creatives, custom audiences, catalogs, signals, tests, activity logs. New entities land paused. | OAuth through Facebook Login for Business inside the client. No developer app. | Meta's pages state no price for the connectors |
| Meta hosted server, your own app or system user (docs) | Same tools | A Meta app with the "Create & manage ads with ads MCP server" use case, then OAuth, a user token or an Employee-role system user token | Meta's pages state no price |
| Hosted third-party connector, for example Pipeboard, Windsor.ai or Madgicx | Pipeboard lists 42 Meta tools; Windsor.ai lists 858 fields and 20 write actions with no delete; Madgicx exposes create, pause and budget writes with manual approval | You log in to the vendor, the vendor holds the Meta token | Pipeboard free plan then $29.90, $99 and $199 a month; Windsor.ai free for one data source then from $19 a month billed annually; Madgicx pricing not on the MCP page |
| Self-hosted open source, for example GoMarble or brijr/meta-mcp | GoMarble: 21 read tools under MIT. brijr: 39 tools on Cloudflare Workers | Your own Meta user token or your own Meta app credentials | Free software, your hosting and your developer time |
Two things on that table are easy to miss. First, every route ends at the same Marketing API data, so no vendor has numbers Meta does not hold; what vendors add is a hosted token, cross-platform coverage, a different set of fields or tools, confirmation habits and support. Second, Meta's server exposes no tool that deletes a campaign, ad set or ad. The ad creation tool page lists ads_delete_custom_audience as its only delete, and the catalog and pixel groups carry deletes for products, feeds and pixel events but none for campaigns, ad sets or ads, while the ads CLI command reference will "Delete an ad campaign and all its child ad sets and ads" and offers --force to "Skip confirmation prompt." The chat route is the gentler one.
Pick the route by who holds the key, not by tool count
Choose Meta's hosted server with your own login when one person's Meta permissions should also be the agent's permissions. A system user or your own app is the answer when the agent has to run without a person, for example a scheduled report. A hosted vendor earns its fee when you need Meta, Google and TikTok in one connector or your team wants someone to call. Self-hosting suits a security team that insists the token never leaves your infrastructure, provided an engineer will maintain it.
With your own login, the agent can do exactly what you can do in Ads Manager, no more, and every change appears under your name in the account history. A vendor asks you to trust its token storage and its definition of a safe write. A system user leaves you owning a long-lived token that someone has to rotate, and that the Meta MCP server only accepts at Employee role, according to the get started page: "Only Employee-role system user access tokens are supported."
A hypothetical pair makes the rule concrete. Company A is an in-house team of three running one ad account, and the PPC lead wants weekly anomaly checks and faster reporting, so the hosted Meta server with her own login inherits her permissions and needs no engineer. Company B is an agency with 40 client accounts across Meta and Google that wants a shared connector for every account manager. Its problem is credential sprawl and client-level write controls, so a hosted vendor with slots per ad account, or Meta's own app route with per-client rules, starts to look cheaper than 40 personal logins. Both answers are right for the question each team asked.
Connect Claude to Meta ads in about ten minutes
Connecting Claude to Meta ads takes a custom connector pointed at https://mcp.facebook.com/ads and a Facebook login. Anthropic's support article says custom connectors are available on every Claude plan, with free users "limited to one custom connector," and that "only Owners and Primary Owners can add them to Team plans."
- In Claude, open Customize, then Connectors. Click "+ Add," then "Add custom connector." On Team or Enterprise, an Owner does this first under Organization settings, then each member connects individually.
- Name it something plain like "Meta ads" and paste
https://mcp.facebook.com/adsas the remote MCP server URL. Click Continue, review the detected authentication settings and click Continue again. - Under Authentication choose "Sign in now." Claude redirects you to the Facebook Login for Business dialog. Sign in with the Facebook account or Meta Managed Account that has access to the ad accounts you want, and grant only the business assets the dialog asks about that the agent should use. Meta's help page states the principle: "Any actions taken on your behalf require your authorisation through the AI agent."
- Start a new chat, press "+", choose Connectors and switch Meta ads on. Ask "List my ad accounts." If the account you expect comes back flagged as not enabled for the ads MCP server, read the failure section below before you touch anything else.
- Ask for a read-only task first, such as spend, CPM and cost per result by campaign for the last 14 days, and check two numbers against Ads Manager. Only when they match should you let the agent create anything, and remember that Meta's tool documentation says write tools "create entities in a paused state; your AI client asks for confirmation before activation."
For Claude Code, the MCP documentation gives the pattern claude mcp add --transport http <name> <url>, so the command is claude mcp add --transport http meta-ads https://mcp.facebook.com/ads, followed by /mcp inside Claude Code to log in through the browser. If you built your own Meta app, Meta's get started page adds --client-id <META_APP_ID> to that command so the OAuth flow uses your app.
ChatGPT needs a Business or Enterprise workspace for write access
Connecting ChatGPT to Meta ads with write actions requires a ChatGPT Business or Enterprise workspace as of November 2026. OpenAI's help center article, updated in late October 2026, says "Custom MCP server connections and full MCP support are available for ChatGPT Business and Enterprise/Edu customers on ChatGPT web," and in its FAQ that "Pro users can connect custom MCP servers with read/fetch permissions." Plus is not mentioned on that page, and custom servers are web only.
- An admin or owner opens Plugins in ChatGPT, selects the plus button and "Add custom MCP server." Enter a name and paste
https://mcp.facebook.com/adsas the public HTTPS endpoint. - Set the authentication method to OAuth so the Facebook login dialog handles the token. If you connect through your own Meta app, Meta's documentation says to enter your app ID as the OAuth client ID under advanced OAuth settings and to set the app's redirect URL to ChatGPT's specification.
- Tick the risk warning, select "Create as a plugin" and review the discovered tools. Test it privately in a new conversation by typing "@" and selecting the plugin. Only admins and owners can publish the connection to the workspace, and on Enterprise they can disable specific actions before publishing.
- Know that ChatGPT keeps a "frozen" snapshot of the server's tools after approval. Meta is still adding tools, so an admin has to refresh the action list before the workspace sees new ones.
A team on ChatGPT Plus that wants write access has to go through a hosted vendor. Madgicx, for example, documents its connector for Plus, Pro, Team, Enterprise and Education plans, with write operations that "require manual approval unless you enable conversation-wide remembering."
Set the ads MCP server rules before the first write
Meta lets anyone with full control of a business portfolio block or cap what agents can do per ad account, and the server enforces those rules whichever client is connected. The help page describes the controls and the July 16, 2026 update to the announcement introduced them. Databricks' October 6, 2026 write-up of the same server puts it plainly: "The server checks every tool call against those rules before it executes."
- Open Settings in Meta Business Suite. Below Integrations, select Ads MCP server. If the entry is missing, your portfolio does not have the feature yet: keep the agent on read-only reporting until it appears, because without these rules the confirmation prompt in your client is the only thing between a typo and a live campaign.
- Select the ad account or catalog to manage.
- Switch each action to Allowed or Blocked. The help page uses two examples: set "Create campaigns" to Blocked and any agent request to create a campaign is denied; set "Edit or set budget" to Blocked to stop all budget changes, or switch "Budgets set above this amount" to Blocked and enter a cap, so that with USD 1,000 entered "requests to set the budget over USD 1,000 will be blocked."
- Confirm the result in the account view, where blocked actions show "Limited" next to "Take actions in this ad account."
We would start every account with campaign creation allowed, activation handled by a person, and a budget cap at roughly the largest single daily budget you run today. Partners can set their own rules on accounts shared with them, and those apply only to the partner's employees, so an agency and its client can each hold a different line. This is the part a CFO will ask about, and it belongs in the same conversation as where humans stay in the loop on ad spend.

The Marketing API access tier only matters if you bring your own app
The Marketing API access tier applies to developer apps, so it affects you only on the own-app route or with a third-party or self-hosted server. Meta renamed the tiers on May 4, 2026 in a developer blog post: "Standard Access" became Limited Access and "Advanced Access" became Full Access, and the qualifying threshold for Full Access "has been lowered from 1,500 to 500 Marketing API calls" in the past 15 days, with an error rate under 15% over the last 500 calls.
The difference between the tiers is quota. Meta's authorization page describes Limited Access as "Heavily rate-limited per ad account. For development only," and warns that "Calls on ANY access level are against production data." The rate limiting reference, which still uses the old names, gives the Ads Management ceiling as "300 + 40 * Number of Active ads" calls per hour at the lower tier and "100000 + 40 * Number of Active ads" at the upper one. A new app with 50 active ads therefore gets about 2,300 management calls an hour, plenty for one analyst and thin for a shared agent across ten accounts.
For the own-app route, Meta's get started page sets the minimum permissions as ads_mcp_management plus either ads_read or ads_management, with catalog_management, business_management, pages_show_list and instagram_basic adding more tools. The permission reference describes ads_management as letting an app "both read and manage the Ads account it owns, or has been granted access to." If you manage other businesses' accounts through your app, expect App Review.
System user tokens follow the steps on Meta's CLI get started page: create a system user under Business Settings, Users, System Users; assign it to the app and to each ad account; click Generate Token and grant the scopes. Keep that token in a secrets manager, never in a chat.
The first month fails in four predictable ways
The most common failure is a connection that works and an account that returns nothing. Jason Poonia of Lucid Media described it on May 4, 2026: the OAuth step succeeds, then the account comes back with a field he reports as is_ads_mcp_enabled: false, because Meta is switching the beta on account by account. The field appears in no Meta document we could open, and we found no published eligibility criteria. If a client account is flagged false, a hosted vendor or a self-hosted server on the ordinary Marketing API is the workaround, and the official connector can stay installed for when the flag flips.
The second failure is trusting the summary over the table. Dewang Mishra of Passionfruit wrote on May 13, 2026 that "Claude sometimes states numbers that do not match what the API actually returned." That is a model problem rather than a Meta problem, and the fix is procedural: ask for the raw rows, have the model compute from them, and spot-check two figures against Ads Manager every time the report format changes. Meta's own reporting tools return more than rows: ads_get_opportunity_score carries recommendations, ads_insights_anomaly_signal flags unusual patterns and ads_insights_industry_benchmark compares you with similar advertisers, and a recommendation from a server still needs a person to agree with it.
The third failure is accidental activation. Every create tool lands paused, but ads_activate_entity exists and the documentation warns that it "starts spending budget." A client that remembers "always allow" for a tool removes the only prompt between a typo and a live campaign. Anthropic's connector article advises against "Allow always" unless you trust the server and the tool, and we would extend that to never allowing activation or budget edits without a prompt, whatever the client.
The last failure is governance drift: system user tokens for departed staff stay valid until someone invalidates them, vendor plans count ad account slots, and a ChatGPT workspace runs on a frozen tool list until someone refreshes it. Put a quarterly check in the calendar that lists every connected client, every system user and every agent rule, and treat it like a user access review. The wider question of which jobs an agent should own at all is the subject of what AI agents can realistically run in a paid media team.
Disconnecting takes a step in the client and a step on Meta's side
Removing the connector stops one client, and removing the grant on Meta's side stops anything else that holds it, so do both.
- Claude: open Customize, then Connectors, and click "Remove" or the three dots next to the connector, as Anthropic's support article describes. Team and Enterprise Owners do this under Organization settings, Connectors.
- ChatGPT: OpenAI's permissions article says "Changing an app permission does not disconnect the app," so open Settings, then Apps or Plugins, choose the app and disconnect the account. An admin can disable the app workspace-wide under Admin Console, Plugins, per the admin controls page.
- Meta, login route: on the Facebook account that signed in, go to Settings and privacy, Settings, Business integrations, select the integration and click Remove, then Remove. Meta's help page says it then "can no longer manage your ads" but "may have information that you've previously shared."
- Meta, system user: Meta's developer docs say "You cannot delete a system user or admin system user, but you can invalidate all access tokens for that user."
When the person whose login the agent uses leaves, remove them from the portfolio under Settings, People. Meta's help page says this "revokes their access to the business portfolio and any business assets," and the Facebook Login for Business documentation says a user token's access "is inherited from your app user's current account access," so the agent loses those assets with them.
Where to start
Connect Meta's hosted server with your own login to one ad account this week, run two weeks of read-only reporting against it, and set the portfolio rules before anyone creates an ad. If the setup, the rules and the reporting routines are more than your team wants to own, that is the kind of work Scalewell sets up for paid teams.
Frequently asked questions
Is the Meta Ads MCP server free?
Meta's announcement, help page and developer documentation state no price for the ads AI connectors as of November 2026, and the product is in open beta. Several vendor blogs describe it as free during the beta, but that is their reading rather than a Meta statement. Third-party connectors publish their own prices, which range from a free tier to a few hundred dollars a month.
Can I connect Meta ads to Claude on the free plan?
The free plan is enough to connect. Anthropic's support article says custom connectors using remote MCP are available to Free, Pro, Max, Ultra, Team and Enterprise users, and free users are limited to one custom connector. Your ad account also needs to be enabled for the beta on Meta's side, which you can check by asking the connected agent to list your ad accounts.
Will using an MCP server get my Meta ad account banned?
Meta's help page says it is safe to connect through its ads MCP server and that bans come from violating advertising standards, such as non-compliant ads or evading enforcement with new accounts. Third-party servers use the same Marketing API your reporting tools use. Keep the ads compliant and the connection itself is not the risk.
Can the agent delete campaigns through the Meta Ads MCP server?
Not as of November 2026. The ad creation and management tool page lists create, update and activate tools for campaigns, ad sets and ads, and the only delete tool there is for custom audiences. The ads CLI is different: it can delete a campaign with all its child ad sets and ads, and its --force flag skips the confirmation prompt.
Do I need a Facebook developer app to use Meta's ads MCP server?
A developer app is optional. Meta's get started page says "Owning a Meta app is not a prerequisite on using the ads MCP server." You need an app only when you want to connect through your own OAuth client, run the agent on a system user token, or build a product for other businesses, in which case the app needs the ads MCP use case and the relevant permissions.