Claude for ads: connect Meta and Google, then let it read and act

"Claude for ads" means connecting your Meta Ads and Google Ads accounts to Claude so it can read live campaign numbers and, when you allow it, act on them: pause a campaign, raise a budget, duplicate an ad. No CSV exports, no screenshots, no pasting tables into a chat window. You ask in plain English and get an answer with the real figures from the account. This guide covers what Claude can and cannot do with ad accounts, how to connect in a few minutes, which prompts are worth your time, and the guardrails to set before you let it change anything.

The three levels of using Claude for ads

Most people who say they use Claude for ads are at level one. The value is at levels two and three.

Level one works for a one-off analysis. It fails as a routine, because the export is the routine. You end up doing the same Friday export you did before, now with a chat window at the end of it.

Level two is where Claude starts paying for itself. The connector gives it tools to list your accounts, pull metrics for a period, break them down by campaign or placement, and read creatives and lead forms. You ask, it queries, you get numbers that match Ads Manager.

Level three is the part that makes people nervous, and it should. Claude pausing a campaign on your behalf is useful on a Tuesday afternoon when CPL doubled and you are stuck in a meeting. It is dangerous if there are no limits. The second half of this guide is about those limits.

What Claude can do once your accounts are connected

The connector exposes two kinds of tools: reading and writing. Reading is always on. Writing is off until you switch it on.

Reading

• List every ad account and Facebook Page you administer, including the ones clients shared with you through their Business Manager.
• Pull Meta Ads metrics for any period: spend, impressions, clicks, CPC, CPM, CTR, reach, leads, cost per result, ROAS, and the rest of the 122 metrics available from the Meta API. Ads that ran on Instagram are included.
• Break those metrics down by placement, device, age, gender, and region.
• Pull Google Ads campaigns with cost, clicks, conversions, and ROAS.
• See organic engagement, Page visits, and follower growth from your Facebook Pages next to the paid numbers.
• List the available fields, so you can ask for a metric by its real name when the friendly name is ambiguous.
• List the creatives and the lead forms attached to an account.

Writing (off by default)

• Pause and reactivate a campaign, ad set, or ad.
• Adjust a daily budget, within a cap you define.
• Create a campaign, ad set, ad, or creative, and upload the media for it. Creation is a separate permission, also off by default.
• Duplicate an ad.
• Pull the change history: who changed what, when, from where, and whether it ran or was refused.

If you only ever use the reading tools, you still get most of the benefit. Plenty of media buyers never turn writing on, and that is a legitimate choice, not a half-measure.

How to connect Meta and Google Ads to Claude

Claude cannot talk to the Meta or Google APIs on its own. It needs a connector that speaks MCP, the protocol Claude uses to call outside tools. If MCP is new to you, What is MCP? A plain-English guide for marketers explains it in five minutes. The short version: MCP is how Claude gets hands.

The setup has three steps.

Step 1: authorize your ad accounts

Sign up for a connector like Vazante and authorize Facebook. That single authorization covers both Meta Ads and Facebook Pages. Vazante then lists every ad account and Page you manage, including the ones inside client Business Manager portfolios shared with you, and you tick the ones to include.

Authorize Google the same way. It brings in your Google Ads accounts, including the ones under a manager account (MCC).

Both go through the standard Facebook and Google login screens. The connector never sees your password. You can revoke access whenever you want, either in the connector's settings (Account and team → Connections → Disconnect) or from your Facebook or Google account. The whole step takes under a minute.

Step 2: add the connector in Claude

In Claude, go to Settings → Connectors → Add custom connector. Paste the address https://vazante.app/api/mcp, then authorize. No API key, no code, no server to run. You need a Claude plan that supports connectors; check Claude's current plan page, because plan features change.

The step-by-step is on the connect Meta Ads to Claude page.

Step 3: ask something you already know the answer to

"How much did account X spend yesterday?" Compare Claude's answer with Ads Manager. Once they match, you trust the pipe, and you can move on to questions you do not know the answer to. Skipping this step is the most common mistake I see: people start with a complex question, get a number that feels off, and blame the tool when the real problem was a date range or a currency setting they never checked.

If you work as a team

Invite the rest of the team by link. Everyone who joins sees the same connected accounts without reauthorizing Facebook or Google themselves. That matters in an agency, where the person who owns the client relationship is rarely the one who set up the Business Manager access. One connection, one set of accounts, one place to look.

The prompts worth using, in priority order

Generic AI content will tell you to "ask Claude to optimize your campaigns." Do not. That prompt has no data behind it, and Claude will give you a generic checklist. The prompts below work because each one maps to a specific query the connector can run.

1. The daily anomaly check

"Compare spend, leads, and CPL for the last 7 days against the previous 7, by campaign. Flag anything that moved more than 30% either way."

This replaces the morning scroll through Ads Manager. Thirty percent is my threshold, not a rule; pick one that fits your account size. The point is that you read a short list of exceptions, not a table of everything.

2. The cheapest lead

"Which campaign had the cheapest lead in the last 7 days, and how many leads did it deliver?"

The second half matters. A campaign with a US$ 4 CPL on 3 leads is noise. A campaign with a US$ 9 CPL on 140 leads is where the budget should go. Ask for volume every time you ask for cost. More on why in CPL: what a good cost per lead looks like.

3. The threshold alert

"Is any ad set running above US$ 20 CPL this week? List them with spend and lead count."

Set the threshold at the CPL that stops making money for you, not at a round number. This is the prompt that finds the ad set that has been quietly burning US$ 60 a day for five days while the campaign average looked fine.

4. The fatigue check

"Show frequency and CTR by ad set for the last 14 days, sorted by frequency, highest first."

When frequency climbs and CTR drops in the same ad set, the audience is tired of the creative. You will see it here days before it shows up in CPL. What to do about it is in Frequency in Meta Ads: when it becomes a problem.

5. The placement breakdown

"Break down last month's spend and cost per result by placement for campaign X."

Ads Manager can do this too, but it takes a handful of clicks and a breakdown menu. Here it is one sentence, and you can follow up with "now by device" or "now by age" without rebuilding anything.

6. The cross-platform view

"Put Meta and Google side by side for the last 30 days: spend, conversions, cost per conversion."

Useful, with one warning. Meta and Google count conversions differently and attribute them on different windows, so the two cost-per-conversion numbers are not the same currency. Treat the comparison as a trend, not a verdict. The full argument is in Meta and Google Ads in one dashboard: the honest way to compare.

7. The client paragraph

"Write three sentences for the client about account X this week: what we spent, what we got, what we are changing. Plain language, no jargon."

Claude is good at this because it has the numbers and the context of the earlier questions in the same chat. It is the last step of a weekly report, not the first. The structure that comes before it is in client reporting for small agencies.

Prompts to skip

• "Optimize my account." No data behind it, no specific query. You get a generic list.
• "What will my CPL be next month?" Claude does not forecast auctions, and neither does anyone else with any accuracy.
• "Which audience should I target?" It can tell you which audience performed. It cannot tell you which one to try next; that is your job.
• Anything that needs CRM data. Claude sees leads as Meta and Google count them. It does not know which ones became customers unless you tell it.

Letting Claude act: the guardrails that matter

Here is where I disagree with most of what is written about AI and ads. The question is not whether Claude should be allowed to touch campaigns. It is what has to be true before it does. Four things.

Writing is off until you turn it on

A connector that can pause campaigns from day one is a liability. In Vazante, write actions are off by default. You turn them on in Account and team → Write actions, and only on the plans that include writing. If you never flip the switch, Claude can read everything and change nothing.

Every action has a ceiling

You define the maximum value a single action can set. Say US$ 500 of daily budget. If you, or Claude misreading you, ask for US$ 5,000, the request is refused before it reaches Meta or Google. The cap protects you from a typo as much as from the model.

Every action asks first

Before any change runs, Claude states what it is about to do and waits for your confirmation in the chat. "Pause campaign Spring Sale, currently at US$ 120 a day. Confirm?" You say yes, or you do not. There is no batch mode where ten things happen while you are not looking.

Every action is logged

Every change is recorded with who asked for it, where it came from (dashboard or connector), the state before and after, and whether it was executed or refused. When a client asks why a campaign was paused on Tuesday, you have the answer in seconds, including the times Claude proposed something and you said no.

Creating campaigns and ads is a separate permission, also off by default. Reading and pausing are one level of trust. Creating something that spends money is another, and it should require a second deliberate choice.

Before you turn writing on: a checklist

• You have used reading for at least a week and the numbers matched Ads Manager every time.
• You know the daily budget above which you would want a human to look. That is your cap.
• Everyone on the team who can trigger actions knows that every action asks for confirmation, and that "yes" in the chat means it runs.
• You have opened the change history once, so you know where to find it when you need it.
• The client, if there is one, knows that some changes may come through the connector and that all of them are logged.

A week with Claude on the account

An illustrative week, with made-up numbers, for a freelancer running six Meta accounts and two Google accounts.

Monday, 8:40 a.m. "Compare last week with the week before, all accounts, by campaign. Flag anything that moved more than 30%." Claude returns four flags. Two are expected: a launch and a paused test. One is a dental clinic where CPL went from US$ 18 to US$ 41. One is a Google campaign that spent 40% less because it hit its daily budget earlier each day.

Monday, 8:52 a.m. "For the dental clinic, break the last 7 days down by ad set with frequency and CTR." One ad set has frequency near 5 and CTR down by half. Found it. "Pause ad set Whitening Retargeting." Claude states the ad set, its current budget, and asks to confirm. Yes. Logged.

Tuesday. Nothing. That is the point. You run the anomaly prompt, it flags nothing new, you move on.

Thursday, 3:15 p.m. A client approves more budget over the phone. "Raise the Meta campaign Fall Promo to US$ 90 a day." Under the cap, one confirmation, done. No login, no navigating to the campaign, no risk of editing the wrong row.

Friday, 4:00 p.m. "For each account, write three sentences for the client: what we spent, what we got, what we changed." Eight paragraphs, one per account. You read them, fix one word, paste them into the emails.

Total time in Ads Manager that week: two visits, both to check a creative preview. Total time typing prompts: maybe twenty minutes.

What Claude gets wrong, and how to keep it honest

Claude with a connector is a fast analyst with no memory of your business unless you give it one. Four failure modes to watch.

It does not know a lead from a customer. The Meta API returns leads as Meta counts them. Claude will happily call a campaign your best performer because it has the lowest CPL, while your sales team knows those leads never answer the phone. Tell it. "Treat the Instant Forms campaign leads as half value; sales says they rarely close." It will adjust its reasoning for that chat.

It compares apples to oranges across platforms. Meta and Google conversions are not the same unit. Claude will put them in one table if you ask. It is on you to read the table as a trend, not a ranking.

It is confident on small samples. Three leads at US$ 4 is not a cheap campaign; it is a coin toss. Add "only include ad sets with at least 20 results" to any ranking prompt, and ask for the count next to every rate.

It answers what you asked, not what you meant. "Pause the expensive one" is ambiguous when two campaigns are expensive in different ways. Name the campaign. The confirmation step catches most of this, but the habit of naming things precisely is what makes the confirmation step boring, which is what you want.

One more thing. A connector complements Ads Manager; it does not replace it. Creative previews, audience building, pixel diagnostics, and policy appeals still live in Meta's and Google's own interfaces. Claude handles the reading and the routine changes. You still open Ads Manager, just less often and with a specific reason.

Claude with a connector vs. Ads Manager vs. reporting tools

People ask whether this replaces the reporting tool they already pay for. It depends on what the tool was doing for you.

Reporting tools such as Supermetrics, Funnel, and Looker Studio are built to move data into spreadsheets and dashboards and keep it there on a schedule. That is a real need, and they do it well. They are readers, not actors; executing changes in campaigns is structurally outside what they do. If your clients want a scheduled visual dashboard, you may keep one. If what you were really using the tool for was to answer "what happened this week and why," Claude with a connector answers that faster, and the dashboard becomes optional. The wider comparison is in Supermetrics alternatives: 6 tools compared.

Who should set this up, and who should wait

Set it up now if you manage three or more ad accounts and spend more than an hour a week reading them. The daily anomaly prompt alone pays back the setup in the first week. Freelancers with a handful of clients and small agencies with a shared team are the sweet spot, because everyone sees the same connected accounts without reauthorizing.

Set it up, reading only if you are in-house on one or two accounts. The questions are still faster than Ads Manager, and you can leave writing off forever without losing anything you needed.

Wait if you spend under US$ 500 a month on a single account and check it once a week. Ads Manager is enough for that. Come back when the account grows or a second one shows up.

Where to start this week

Connect one account, reading only. Run the anomaly prompt every morning for two weeks and compare it with what you would have caught on your own. If it finds things you missed, or finds them earlier, add the other accounts. Then, and only then, decide whether to turn writing on, starting with a low cap.

Vazante's Basic plan starts at US$ 23 a month billed annually, with a 7-day free trial and no credit card, and it includes the Claude connector in read-only mode. Writing is on the plans above it. The details are on the pricing page. Set up your first prompt before you close this tab; the connection takes less time than reading this section did.

Two neighboring setups worth knowing: how to use ChatGPT for marketing, if your team already lives there, and connect Instagram to Claude, to read organic next to paid.