Meta Ads with Claude: From CSV Audits to the Official Connector

Most people who ask me about Claude and Meta ads want the same thing: an assistant that looks at the account, says what to cut and what to scale, and writes the next round of creative. That is realistic today. The part people get wrong is the order. Start with an export and a good prompt, learn to trust the analysis, and only then give Claude a live connection to an account that spends real money.

Step 1: Export the right data

In Ads Manager, set the date range to the last 30 days, choose the ad level, and export the table as a CSV. Before you export, customize the columns so the file has what an analysis actually needs:

• Campaign, ad set and ad name, plus delivery status
• Amount spent, impressions, reach and frequency
• Link clicks, CTR and CPM
• Results, cost per result, and purchase ROAS if you sell online

Then export the previous 30 days the same way. Two periods side by side tell you far more than one, because a number is only good or bad compared with something.

Name your ads consistently before any of this. If your ad names say what the hook, format and audience are, Claude can find patterns across the account. If they say "Ad 3 copy final v2", it cannot.

Step 2: Give Claude your own targets

Many ad prompts floating around include industry benchmarks for CTR or ROAS. I would leave them out. A 2x ROAS can be a great result for one business and a loss for another. What matters is your margin and what a customer is worth to you.

You are reviewing a Meta ads account for [business].
Context:
- Goal: [purchases / leads / signups]
- Target cost per result: [amount]. Break-even point: [amount].
- Average order value or lead value: [amount]
- Anything unusual this month: [sale, stockout, new creative, none]

Attached are two CSV exports: last 30 days and the 30 days before.
1. Summarize what changed between the two periods in 5 lines.
2. List ads to pause, with the reason based on my targets.
3. List ads to scale, with a suggested budget change and the risk.
4. List ads with too little data to judge. Do not guess on these.
5. Find patterns in hooks, formats or audiences, using the ad names.
6. Give me three actions for the next 7 days, ranked by impact.
Show the numbers behind every recommendation.

Point 4 is the one I care about most. An ad with a small number of results can look like a winner or a disaster purely by chance. A good analysis says "not enough data yet" out loud.

Step 3: Write variants from what is working

Once you know which ad is carrying the account, use it as the control and write a few focused variants. The mistake is generating thirty at once. You will not have the budget to test them properly, and the results will be noise.

This is my best performing ad: [paste primary text, headline, and describe the visual].
It works for [audience] because [your guess].
Write 6 variants:
- 3 that change only the opening line
- 3 that change only the angle (for example pain, outcome, proof)
Keep the offer, facts and brand voice identical.
Do not invent numbers, testimonials or guarantees.
Label each one so I can track it in the ad name.

Test two or three against the control at a time. When one wins, it becomes the new control and you repeat.

For competitor research, browse the Meta Ad Library in your browser, take screenshots of the ads that have been running the longest, and give them to Claude. Ask what angles they share and which ones nobody in your category is using. Do not copy their ads. Use them to find the gap.

Step 4: Connect Claude directly with Meta's official connector

When the export workflow feels routine, you can remove the export step. Meta launched Meta ads AI connectors in open beta in April 2026. The core of it is an official MCP server that Meta hosts at https://mcp.facebook.com/ads, with tools for reporting, campaign and ad management, catalogs and signals. The developer overview lists what it can do.

To use it in the Claude app, add it as a custom connector with that URL, then sign in with Facebook and choose the business portfolio you want to connect. If you log in to Claude Code with the same Claude account, connectors from the app are available there too. Meta's get started page also covers connecting through your own Meta developer app, and Meta offers a command line tool as another route.

A few things make this safer than it sounds. According to Meta's docs, write tools create campaigns, ad sets and ads in a paused state, and your AI client asks for confirmation before anything is activated. Since July 2026, people with full control of a business portfolio can also set rules that limit what AI agents may do on the ad account. Use both.

Start with read-only questions for the first week:

Show spend, results and cost per result by campaign for the last 7 days.
Which ad sets have frequency above 3 this week?
Compare this week with last week and tell me what changed most.

Only move to changes after you have checked a few answers against Ads Manager yourself.

Optional: a skill pack for audits

If you run ads for several accounts, claude-ads is an MIT licensed skill pack for Claude Code that runs structured audits across Meta and other platforms from exports or account reads. It is read-only by default. The README shows the current install, including the plugin route:

/plugin marketplace add agricidaniel/claude-ads
/plugin install claude-ads@ai-marketing-hub-claude-ads

Read what a pack does before you install it, and check the README for current commands, since they change between versions.

The rule for anything that spends money

Claude suggests. You review. You approve. Then the change happens. That holds for budget changes, launches and pauses, whether the change goes through the connector or by your own hand in Ads Manager. I use Claude every day, and I still would not let any assistant move budget without me reading the plan first.

This week, export two periods, run the audit prompt with your real targets, and compare its pause list with your own instinct. Where you disagree, ask Claude to show its numbers. That conversation will teach you more about your account than any benchmark.