Meta Ads MCP: the official server, the alternatives and a safe setup

A Meta Ads MCP server is a connector that lets an AI assistant such as Claude, ChatGPT or Codex read and change your Meta ad account through the Model Context Protocol. You can get one in three ways: Meta’s official hosted server, a hosted third-party server, or an open-source server you run yourself with your own Meta credentials.

Meta’s own server has been in open beta since April 2026, and it’s the only option below whose documentation also covers catalogs, signal health, A/B tests and conversion lift studies. The others win on other things: more ad platforms, no write access at all, or full control of the code. And the servers that only wrap the Meta API report conversions as Meta counts them, which isn’t the same as the leads you received.

Meta Ads MCP servers compared

Five options, each described from its own documentation as of November 4, 2026:

Before an agent touches live budget, two columns matter more than the rest: what it can change, and how it authenticates.

Meta’s official ads MCP server

Meta announced Meta ads AI connectors in open beta on April 29, 2026: an ads MCP server and an ads command line interface (CLI), “built for businesses of all sizes across regions”. The announcement says that for MCP, no developer credentials, API setup or coding are required.

According to Meta’s ads MCP server overview, the server is Meta-hosted at https://mcp.facebook.com/ads and “works with any MCP-compatible AI agent”. Its tools come in seven groups:

• reporting: spend, impressions, CTR, CPC, CPM and conversions for campaigns, ad sets and ads, plus trends, anomalies and industry benchmarks;
• ad creation and management: campaigns, ad sets, ads, creatives, custom audiences, Instagram boosts and Ad Library searches;
• catalogs: catalogs, product data and feed troubleshooting;
• signals and datasets: signal health and quality;
• help: Meta Business Help Center searches;
• A/B tests and conversion lift studies: create, manage and read them;
• activity logs: changes on the ad account, like the campaign history in Ads Manager.

The write rules are spelled out in Meta’s ad creation tools reference: “Write tools create entities in a paused state; your AI client asks for confirmation before activation.” Activation is its own tool, the one that “starts spending budget”; another updates fields on existing campaigns, ad sets and ads. The reference lists no tool to delete campaigns, ad sets or ads; its only delete removes a custom audience.

On July 16, 2026, Meta added two things. You can now connect any AI application through your own developer app, and new “ads MCP server rules” give anyone with full control of a business portfolio “the ability to govern what AI agents can do on their ad account, from budget changes to catalog updates.”

Sign-in goes through Facebook Login for Business (OAuth), with a Facebook account or a Meta Managed Account. Meta’s get started guide sets the minimum at ads_mcp_management plus either ads_read or ads_management.

Meta writes it for advertisers who’d rather manage ads “through a chat interface instead of Ads Manager or code”, and for developers building AI-native ad workflows. If you need breadth on Meta, start here. What it can’t give you is a number Meta doesn’t have: the conversions it reports are Meta’s attributed count, which drifts from the leads you actually received (here’s why the two CPLs disagree).

Third-party and open-source Meta Ads MCP servers

Three GitHub projects cover the other two routes (READMEs and code checked on November 4, 2026).

Pipeboard Meta Ads MCP (about 1.3k stars). The README calls it “an independent open-source project that uses Meta’s public APIs” and recommends the hosted version at https://meta-ads.mcp.pipeboard.co/: you log in to Pipeboard and connect your Facebook ad account, or append a Pipeboard API token to the URL. Running it yourself means creating your own Meta developer app. It’s one of five Pipeboard servers (Meta, Google, TikTok, Snap and Reddit). Its create tools default to paused, and the README promises “explicit confirmation on every write”. The code is under the Business Source License 1.1. The README itself calls Meta’s official connector “the safest single-platform option” and pitches Pipeboard as the cross-platform choice, which is a fair summary.

GoMarble Facebook Ads MCP server (MIT license). A Python server that runs on your machine. Its 21 tools read accounts, campaigns, ad sets, ads, creatives, insights and change history, and none of them write. You pass a Meta access token (the README suggests one with ads_read) as a launch argument in your client’s config, or let GoMarble’s installer create one, which is then saved locally. When the token expires, you generate a new one and update the config.

meta-mcp by brijr. A Cloudflare Worker “intended to be deployed in your own Cloudflare account with your own Meta app credentials”. It’s built for products rather than for one person: every MCP request needs a JWT issued by your own auth system, and Meta tokens are stored encrypted per workspace. Its 39 tools create, update and delete campaigns and ad sets, upload media, manage audiences and run bulk changes; in the code, new objects default to paused. The repository has no license file, so ask the author before building on it.

Other hosted Meta Ads MCP products exist. Judge any of them on the table’s questions: what it reads, what it can change, how it signs you in, and where your token ends up.

Meta’s numbers tell an agent what Meta counted; DropCPL gives it the CPL of every creative × landing page pair, counted on the leads you actually received. Connect it next to whichever Meta Ads MCP server you pick and ask both in the same chat. Start free →

How to connect Claude, ChatGPT and Codex

Every remote MCP server connects the same way: you give the client a URL, a browser window opens so you can sign in (OAuth), and the tools show up in the chat. For DropCPL’s exact steps per client, see connecting AI assistants to DropCPL.

Claude

In Claude.ai and Claude Desktop, a remote server is a custom connector. Anthropic’s help article on custom connectors says they’re available on the Free, Pro, Max, Ultra, Team and Enterprise plans, with Free limited to one custom connector. On Team and Enterprise, an owner adds the connector for the organization and members then connect individually. In Claude Code it’s one command, claude mcp add --transport http <name> <url>, followed by /mcp to sign in. For Meta’s server, Meta’s guide adds --client-id <META_APP_ID>, using your own Meta app.

ChatGPT

OpenAI’s current guide adds a remote server from Plugins in ChatGPT: choose Add custom MCP server, enter the URL, configure authentication, accept the risk warning and create it as a plugin. Account and workspace policies apply, so on a business workspace an admin may need to allow it. For Meta’s server, Meta’s guide picks OAuth and enters your Meta developer app ID as the OAuth client ID.

Codex

Codex takes codex mcp add <name> --url <server-url>, then codex mcp login <name> for the OAuth sign-in. The configuration lives in ~/.codex/config.toml and, per OpenAI’s Codex MCP docs, it’s shared by the Codex CLI, the IDE extension and the ChatGPT desktop app. If a server expects a pre-registered OAuth client, as Meta’s does with your app ID, Codex accepts one with --oauth-client-id. One setting is worth adding for any ads server:

[mcp_servers.dropcpl]
url = "https://mcp.dropcpl.com/mcp"
default_tools_approval_mode = "writes"

With writes, Codex runs the tools a server marks as read-only and asks you before anything else.

Is it safe to let an AI agent run your Meta ads?

Treat the agent like a new media buyer on day one: it can look at everything, and it needs sign-off before anything spends.

Start read-only. In Meta’s permissions reference, ads_read lets an app pull ads report information, while ads_management lets it “both read and manage” the ad account. Grant the first until you trust the setup. DropCPL’s consent screen defaults to read-only and makes write access an explicit opt-in, for one client subaccount. GoMarble’s server has no write tools at all.

Paused by default, confirmation before spend. An agent that misreads “set up a test” can build a whole campaign structure. If everything it creates is paused, that mistake costs nothing until a person activates it. Meta’s server and DropCPL always create new objects paused and make activation a separate step, which holds whatever your client’s approval settings are. Pipeboard and meta-mcp default to paused, but their create tools accept a status, so an agent could ask for active. The confirmation before activation is the second lock, and your client is the one that shows it. DropCPL flags switching an ad on or off as a destructive tool.

Don’t paste long-lived tokens. With OAuth servers (Meta’s, hosted Pipeboard, DropCPL) you never paste a token: the client gets one through the browser sign-in, and DropCPL’s access tokens expire on their own. Servers that take a raw token put it somewhere it can leak: GoMarble’s config passes a Meta access token as a launch argument, and Pipeboard offers an optional token-in-URL format. Config files get synced and URLs end up in logs, so prefer OAuth. If you must use a token, make it one you can revoke without breaking anything else.

Check the approval settings in your client. Claude asks before calling a connector’s tools in normal chats, and Anthropic advises clicking “Allow always” only for servers you trust; in Research, Claude can call connector tools “automatically without further approval”. Anthropic also warns that a malicious server can carry hidden instructions, so connect only servers from organizations you trust. In Codex, the writes approval mode shown above does the same job.

Read the log afterwards. Meta’s activity logs tool shows what changed on the ad account, and DropCPL records every write action in its Action Log. If several people connect agents to the same business portfolio, Meta’s ads MCP server rules are where you decide what those agents may touch.

What to ask once it’s connected

Good prompts name the account or campaign, the date range and the metric, and they ask for a plan before any write. Some example prompts for media buyers (the amounts are examples), with the servers that can answer them:

On DropCPL, ask for edits as patches (“change this headline to…”) rather than rewrites: the agent then replaces only that exact text instead of regenerating the whole landing page. If you’re running a structured creative test, hand the agent your rules first; a creative testing framework for Meta ads gives it spend thresholds and a point at which to call a winner.

Where DropCPL fits (and where it doesn’t)

DropCPL’s MCP server covers a deliberately narrow slice of the Meta API. It exposes the same tools as the in-app AI agent, and they’re built around one question: what does each creative × landing page pair really cost per lead?

What it adds:

• First-party CPL per creative × landing page. Clicks are logged on a tracking domain that is the landing page’s own domain, leads are counted by DropCPL, and spend per ad is synced from the Meta Ads API. The agent reads creative attribution on the leads you received, not on Meta’s modelled count, and can check whether a winning landing page holds across ads (how to read that matrix).
• Landing pages next to the ad tools. The agent can create a landing page, edit it, validate it, publish it and check the live URL, then launch a creative as a paused ad that points at it.
• Meta ad tools for testing. Link or create a campaign, add a creative from a public URL, create an ad set, launch ads paused, switch an ad on or off, sync spend.

Where something else is better:

• Full Ads Manager coverage. For edits on live campaigns and ad sets, budget changes included, plus audiences, catalogs, A/B tests and lift studies, use Meta’s official server. DropCPL leaves budget and targeting edits to Ads Manager on purpose, and it switches ads, not campaigns or ad sets: the campaigns and ad sets it creates are activated on Meta’s side.
• More than one ad platform. DropCPL is Meta only. Pipeboard also covers Google, TikTok, Snap and Reddit.
• Reporting with zero write access. GoMarble’s server can’t write at all; Meta’s server with ads_read and DropCPL’s read-only scope are the other ways to keep an agent hands-off.

For a lead gen account, the setup that works is both: Meta’s server for account operations, DropCPL for CPL per creative × landing page and for the landing pages themselves, in the same conversation. Claude’s Free plan won’t do it, since it allows one custom connector. If you’re still building the funnel, start with Facebook ads for lead generation.