Meta Ads MCP Plus Your CRM: Which Campaigns Made Revenue

Meta's server details are as of October 2026, in open beta. The example figures below are made up and labelled; the arithmetic is real.

The split is simple and it is worth keeping strict, because it decides which connector Claude calls for which half of a question.

So the revenue truth comes from one side and the levers sit on the other, and Claude is the place where the two meet. The decision stays with you.

1. Meta. In Claude, open Connectors in settings, choose Add custom connector, paste https://mcp.facebook.com/ads, connect and sign in with Facebook Login for Business. Approve ads_mcp_management and ads_read first. The full walkthrough is in Meta Ads MCP: setting up Meta's official server.
2. LeadJourney. In LeadJourney, open Ask Atlas, then Settings, then the MCP tab, copy the server URL (https://api.leadjourney.io/mcp), add it to Claude the same way and sign in. The workspace it reaches is decided at sign-in. See LeadJourney in Claude.
3. Check both. Ask one question per connector: "From Meta, list my ad accounts" and "From LeadJourney, closed revenue by Meta campaign last month."

Two plan details. Claude's Free plan allows one custom connector, and this setup needs two, so it runs on Pro, Max, Team or Enterprise. And LeadJourney's MCP connection is included from the Growth plan, or a €49 a month add-on below it (€99 a month flat for agencies). ChatGPT works too; see LeadJourney in ChatGPT and the Meta section of the setup post.

An example B2B account with one lead generation campaign and four ad sets. The question: where should the budget go? First, what Claude gets from Meta's server alone, over the 90 days that ended a month ago (so most deals from those leads have had time to close):

On Meta's numbers the interest stack is the winner and the lookalike is expensive. Now the same ad sets with LeadJourney's CRM outcomes, pulled in the same conversation:

The ranking flips. The cheapest leads came from the ad set that closed nothing, and the "expensive" lookalike produced more than half the revenue on about a quarter of the spend. An assistant reading Meta alone would have moved budget the wrong way, confidently.

The proposal Claude should return

With current daily budgets read from Meta, ask for a proposal that keeps the total level and moves money in a modest step:

Note what the proposal does not do: it does not switch the interest stack off on six qualified leads and no deals. That is a small sample, and the next quarter can show something different. It moves a step, keeps the experiment running, and comes back to it.

The same eight steps every week or two. Steps 1 to 4 are reads and cost nothing; steps 5 to 7 are where a person decides.

1. Pick a mature cohort. A period whose leads have had your usual time to close, not last week.
2. Revenue from LeadJourney. Qualified leads, booked demos, deals won and closed revenue per Meta campaign and ad set for that cohort, under the attribution model you report on.
3. State from Meta. Current daily budgets, status, optimisation goal and spend for the same ad sets, through ads_get_ad_entities.
4. Join and flag. Claude matches the two by name and lists unmatched rows and ad sets with too few qualified leads to judge.
5. Propose. A table of current and proposed budgets with one line of reason each, total spend stated before and after.
6. Approve. A person reads the table, edits it, and says which rows to apply.
7. Apply. In Ads Manager by hand, or through Meta's ads_update_entity after the explicit approval in the conversation. Meta's pages do not list every field that tool changes, so check the result in Ads Manager either way.
8. Record. Ask Meta's server for the activity log of the session, and save the proposal table as the note of what changed and why.

LeadJourney can save step 2 as a report, so the next run starts from the same definition rather than a prompt typed slightly differently. Ask: "Save this as a report called Meta ad sets by closed revenue."

Name the connector in each prompt, so Claude does not try to answer a revenue question from Meta or a budget question from LeadJourney.

1. "From LeadJourney: for Meta ad sets with clicks between [date] and [date], show qualified leads, deals won, closed revenue and cost per qualified lead, first-click model."
2. "From Meta: for the same ad sets, current daily budget, status, optimisation goal and spend in that period."
3. "Join the two by ad set name. List anything you could not match, and mark ad sets with fewer than 10 qualified leads as too small to judge."
4. "Propose new daily budgets that keep the total the same and move no ad set by more than a modest step. Show current, proposed, change and a one line reason. Do not change anything."
5. "Apply rows 2 and 4 only." (or apply them yourself in Ads Manager)
6. "From Meta: show the activity log for this account for the last hour."

More single-connector prompts for reporting, audits and creative review are in Claude for Meta Ads.

The same pair works for building, not only for budgets. LeadJourney says which audience, offer and landing page produced deals; Meta's server builds a draft that copies it. With ads_management approved:

• "From LeadJourney: which landing pages and Meta ads had the highest closed revenue per lead last quarter?"
• "From Meta: create a paused campaign for [new market] with one ad set copying the targeting and optimisation goal of Lookalike customers, a daily budget of €30, and an ad using the creative of our top ad by closed revenue. Do not activate."

Meta's server creates campaigns, ad sets and ads paused, and spend starts only through ads_activate_entity. Review the draft in Ads Manager and activate it yourself. Meta documents no dry run and no tool to delete a campaign, ad set or ad, so a draft you do not want has to be removed in Ads Manager.

• Proposal first, alwaysClaude returns a table; a person approves rows. Write "do not change anything" into every analysis prompt until the approval step.
• Read permissions by defaultRun the weekly analysis on ads_read. Add ads_management for the sessions where you build drafts or apply an approved table.
• LeadJourney cannot act on MetaIts MCP server reads, and changes only reports and settings inside LeadJourney. Nothing in Meta moves because of it, and it runs only when you send a message.
• Small steps on live ad setsMeta gives no official percentage for budget changes and the learning phase, only that a tiny change is unlikely to restart learning and a tenfold one may. The popular 20% rule is not Meta's. Move in steps and watch the ad set.
• Mature cohorts onlyJudge on leads old enough to have closed. Last week's leads have no revenue yet, and an ad set judged on them always looks bad.
• Enough data to judgeSet a minimum of qualified leads before an ad set is ranked at all, and say the attribution model next to every revenue figure.
• Check what happenedRead Meta's activity log after every session that changed something. It is the record of what the assistant actually did.

If you own the business portfolio, also look at the ads MCP server rules Meta added in July 2026, which let a portfolio owner govern what agents may do, from budget changes to catalog updates. Meta had not published a documentation page on them when we checked, so look for them in your business settings.

Some disagreement between the two connectors is by design. Know which before you call a number wrong.

For the full picture of why platform numbers and CRM numbers rarely match, see why GA4, Meta and Google conversions don't match your CRM.

The weekly workflow fixes where the budget goes. It does not fix what Meta's algorithm optimises for inside each ad set, which is still the form fill. That is a separate job: sending the CRM stages, such as a qualified lead or a won deal, back to Meta as conversion events through the Conversions API (offline conversion tracking), so delivery learns from the leads that closed.

LeadJourney does that for the stages you map, as conversion events with the fbclid and hashed match keys. It sends events; it does not change campaigns, budgets, targeting or ads. How to choose the stage is in send CRM data back to Meta and Google and Meta Conversions API for lead generation. Once qualified leads arrive in Meta as an event, Meta's own reporting through the MCP server starts showing them too, and Meta's automated rules can use them.

LeadJourney is one script on your site or in your GTM container, live in about 21 minutes, tracking every click, form fill and call server-side and first-party at 95%+ accuracy. It keeps the fbclid on the visitor, joins the lead to the deal in HubSpot, Salesforce, Pipedrive, Close, Attio, GoHighLevel, ActiveCampaign, Odoo, Zoho CRM or Microsoft Dynamics 365, and imports Meta's spend, so closed revenue, cost per qualified lead and ROAS sit on every Meta campaign, ad set and ad.

The MCP server serves those rows to Claude, ChatGPT and other MCP clients from one endpoint with OAuth. It reads journeys, spend and CRM outcomes; it changes only reports, dashboards and attribution settings inside LeadJourney; it cannot pause a campaign, move budget, create an ad or edit a CRM record. Inside the dashboard and in Slack, Atlas, LeadJourney's built-in AI analyst, answers the same questions without a second connector.

• Meta Ads MCP: setting up Meta's official server
• Claude for Meta Ads: connector, skills and 25 prompts
• Why your ad platform's MCP gives Claude the wrong CPL
• Automated budget rules on cost per qualified lead
• Ads MCP servers compared: Meta, Google, LinkedIn, TikTok and more
• MCP vs API vs Zapier for ad reporting
• Which ad copy attracts leads that close
• See lead quality per channel and campaign
• The LeadJourney MCP server
• LeadJourney in Claude