Meta Just Killed the $10K Ad Agency: How to Connect Claude to Your Ad Account (Free Setup Guide)

TL;DR: Meta opened an official connector called Meta Ads AI Connectors in beta on April 29, 2026, an MCP server at mcp.facebook.com/ads that lets any MCP-compatible agent, including Claude Code, read and manage your ad account directly. On July 16, 2026, Meta made developer access self-serve (no review queue) and shipped permission rules so business admins control exactly what an agent can touch. Run one command, and Claude can summarize spend, launch paused campaigns, and run A/B tests in plain English. Stack the free, open-source claude-ads skill on top and it audits your account and scores it before anything spends. This guide covers setup, auth, safety, and the exact command, no guaranteed results, just the real steps.

What changed on July 16, 2026

Meta’s Ads AI Connectors have been in open beta since April 29, 2026. The official MCP server lives at mcp.facebook.com/ads and works with any MCP-compatible agent, not just Claude. What changed this week is access.

Before July 16, getting your developer app approved for the ads MCP server use case meant sitting in Meta’s review queue. As of this week, that queue is gone for this specific use case. You create or reuse a developer app at developers.facebook.com/apps, add the “Create & manage ads with ads MCP server” use case, and you are in.

Meta also shipped what it calls ads MCP server rules. Business portfolio admins can now restrict what an AI agent is allowed to do inside a connected account, budget changes, catalog updates, and more, before any agent gets near the account. That is the piece that makes this safe to actually use rather than just interesting to read about.

Why this matters if you have ever paid an ad agency

A lot of small and mid-size ad accounts pay somewhere around ten thousand dollars a month for an agency to do three things: check performance, adjust budgets, and report back. Those three things are now something you can ask an AI agent to do directly, in plain English, against your own account, for the cost of a developer app and a free skill install.

That does not mean an agency adds zero value. Strategy, creative direction, and judgment calls still matter. But the mechanical layer, the “read the account, tell me what is happening, make the obvious fix” layer, just became something you can run yourself.

Step 1: Create your Meta developer app

Go to developers.facebook.com/apps and create a free developer app, or reuse an existing one. Add the use case labeled “Create & manage ads with ads MCP server.” As of the July 16 update, this is self-serve. No review queue, no waiting on Meta’s team to approve you before you can connect.

Copy your App ID. You will need it in the next step.

Step 2: Run the one command

This is the exact command from Meta’s own Get Started documentation. Meta documents Claude Code as a first-class client for this connector.

claude mcp add --transport http --client-id <META_APP_ID> meta-ads https://mcp.facebook.com/ads

Replace <META_APP_ID> with the App ID from Step 1. Claude Code prompts you to complete login on first use, and from that point forward your ad account is wired into whatever agent you are running.

New campaigns created through the connector arrive paused by default. That is a real safety detail worth knowing before you use it: nothing spends without you approving it first. Budget or targeting changes made to campaigns that are already active do apply immediately, so review any live-campaign edit before you approve it.

Step 3: Choose your auth path

Meta’s Get Started doc, updated July 14, 2026, supports two authentication methods:

• OAuth via Facebook Login for Business. No manual tokens. This is the standard path for interactive, day-to-day use inside Claude Code.
• A user access token, generated through Graph API Explorer or your own OAuth flow. This is what you need if you want a scheduled agent to run headless, with your laptop closed.

Either path requires the same seven permissions: ads_mcp_management, ads_read, ads_management, catalog_management, business_management, pages_show_list, and instagram_basic.

If you want to run this programmatically, without an interactive Claude Code session, you POST directly to https://mcp.facebook.com/ads with an Authorization: Bearer <ACCESS_TOKEN> header, using JSON-RPC. That is the mechanism that makes a scheduled, nightly agent possible (see Step 6 below).

What the connector can actually do

Meta groups the connector’s capabilities into seven tool categories:

1. Comprehensive reporting: pull performance data across campaigns, ad sets, and ads.
2. Ad creation and management: build and edit campaigns, ad sets, and ads.
3. Catalog creation and management: manage product catalogs for dynamic ads.
4. Signals and datasets: work with pixel and conversion API data.
5. Help and troubleshooting: surface Meta’s own documentation and fixes.
6. A/B tests and conversion lift studies: set up and read experiment results.
7. Activity logs: see exactly what changed on the account and when.

In practice, this means you can ask a plain-English question like “summarize my top-spending campaigns this week” and get a real answer pulled from your account, or ask it to launch a new campaign, and it arrives paused for your review.

Step 4: Set the safety rules first

Before you hand this to anyone else on your team, or before you connect a business portfolio that includes client ad accounts, go into your Business Portfolio settings and configure the ads MCP server rules. This is where you restrict what an agent is allowed to touch, budget changes, catalog edits, and other sensitive actions, at the account level, independent of any single Claude Code session.

Two rules worth following regardless of your setup:

• Use official Meta paths only. Authenticate through developers.facebook.com or Meta’s own OAuth flow. There is no legitimate shortcut around this.
• Review access scope before connecting a shared portfolio. If your business portfolio includes client ad accounts, a connected app can see all of them by default. Check what is actually in scope before you connect anything.

Step 5: Stack the free audit skill on top

The connector by itself gives you a chat interface to your ad account. What turns it into something closer to an actual media buyer is stacking a free, open-source skill called claude-ads on top of it, built by Agrici Daniel and released under the MIT license at github.com/AgriciDaniel/claude-ads.

It runs more than 250 checks against your connected account and returns a weighted 0-100 audit score, flagging wasted spend and structural issues. It is read-only by default: it tells you what is wrong, and any live change it proposes is gated behind your explicit approval. Nothing gets fixed automatically without you saying yes.

This is Cooper’s own framing, not Meta’s or the skill author’s: stacking the official connector with a free, independent audit skill on top is what turns a read-and-chat tool into something that functions like a full media buyer, minus the ten-thousand-dollar invoice.

Step 6: Automate the boring part

Because auth supports a Bearer token, you can point a scheduled agent at the connector and have it run without you present. A simple, high-value version of this: a nightly wasted-spend check.

/schedule daily at 6am: run a claude-ads wasted-spend audit on my Meta ad account, summarize anything over the threshold, do not make any changes

Keep the schedule read-only to start. Once you trust the flags it surfaces, you can decide which specific actions, if any, you want it to take automatically, and which should always come back to you for approval.

The rollout caveat nobody mentions

Meta is still enabling ad accounts in waves, and that gate is enforced per ad account on Meta’s side, not per Claude account and not tied to which setup path you used. Setup and login will work either way. But if you have gone through every step above and your ad account errors on data access after connecting, it is not something you set up wrong. There is no request form for this. Waiting is the fix, so give it a few days and recheck.

What this does not do

This setup helps you see what is happening in your ad account and catch obvious waste faster than logging into Ads Manager manually. It is not a guarantee of a specific ROAS, revenue outcome, or performance improvement, and nothing about connecting Claude to your ad account changes how Meta’s ad auction works. Treat the audit score as a diagnostic you act on, not a promise.

FAQ

Is Meta’s ads MCP connector official? Yes. It is Meta’s own product, announced and documented at developers.facebook.com, running on the endpoint mcp.facebook.com/ads. It is not a third-party workaround.

Do I need developer experience to set this up? No. Creating a developer app and adding a use case is a form-fill process at developers.facebook.com/apps. The one command you run afterward is copy-paste, with your App ID dropped in.

Will this spend my budget without me approving it? New campaigns created through the connector start paused by default. Budget or targeting edits to campaigns that are already active do apply immediately, so review those before approving.

What is claude-ads and is it made by Meta? No. claude-ads is a free, MIT-licensed, independent project by Agrici Daniel on GitHub. It is not affiliated with Meta. It adds an audit and scoring layer on top of the official connector.

Can I let this run without me watching it? Yes, using a Bearer token and the scheduling command shown above, but start read-only. Give it write access only for specific, already-reviewed actions.

What if I manage client ad accounts? Review the ads MCP server rules and access scope in your Business Portfolio settings before connecting anything. A connected app can see every account in scope unless you restrict it.