What Meta Ads MCP Can and Can't Do (July 2026)

Updated July 2026.

Meta's official Ads MCP server, launched April 29, 2026 at mcp.facebook.com/ads, gives AI assistants like Claude and ChatGPT 29 tools for reporting on and managing Meta ad accounts. It can pull performance data, build campaigns (which land paused), edit live objects, and manage product catalogs — but it has no memory between sessions, no continuous monitoring, and no change history. Here's the verified inventory of both halves.

What the Meta Ads MCP can do

Per Meta's launch announcement and the Business Help Center, the server exposes 29 tools spanning three officially described areas — campaign management, catalog management, and reporting — with community documentation adding assets and diagnostics as further groupings. Concretely:

• Reporting and insights. Pull performance for any date range, break results down by placement, audience, or creative, compare periods, and surface anomalies — all in plain English, no GAQL-style query language required.
• Campaign management. Create campaigns, ad sets, and ads from natural language, and edit existing ones: budgets, targeting, status. Everything the MCP creates starts paused and must be activated manually in Ads Manager.
• Catalog management. Create product catalogs, add product data, and troubleshoot data feed issues — a genuinely underrated capability for ecommerce accounts.
• Diagnostics. Tools for checking account and signal health, per community tool inventories of the beta.

Setup is deliberately simple: paste the URL into a paid AI client that supports remote connectors, authorize with your own Meta login (OAuth, no API keys, no developer app), and you're connected — assuming your account has cleared Meta's gradual rollout. If you hit is_ads_mcp_enabled: false, see our troubleshooting guide.

Credit where due: for ad-hoc analysis, this is the best interface Meta has ever shipped. "Which ad sets drove CPA up last week, and what changed in their delivery?" is now a 20-second question.

What the Meta Ads MCP can't do

1. No memory between sessions. MCP is a stateless protocol. Close the chat and everything is gone — the analysis, your break-even ROAS, the anomaly it spotted yesterday. Next session, the assistant re-reads the account from scratch and you re-type your context.
2. No continuous monitoring. The assistant only acts when prompted. If CPA doubles at 2 a.m. on Saturday, nothing happens until a human opens a chat and asks.
3. No change history. The MCP keeps no record of what it changed, when, or on whose instruction. Meta's native activity history exists but was never designed to reconstruct AI sessions or serve as an audit trail — a gap we cover in the ad account audit trail.
4. No measured causality. Ask "why did ROAS drop?" and you get a plausible narrative assembled from the data in front of it — not a measured link between a specific change and a specific performance move.
5. No draft layer for edits. New objects land paused, but edits to existing campaigns — budget changes included — go live immediately. There is no staging, review, or approval step in the protocol.
6. Meta only. The connector sees one platform. If you run Google, LinkedIn, or TikTok alongside Meta, each needs its own connector, and blending them in one analysis is manual work.
7. Rate limits bite on large accounts. Standard-access accounts throttle under Meta's Marketing API limits — roughly 200 calls per hour by community measurement. Agents that fan out parallel queries hit this fast.
8. No scheduled output. No recurring reports, no alerts, no digest in Slack every morning. Every output requires a fresh prompt.

Where MCP ends and a persistent layer begins

Job	Meta Ads MCP	Persistent layer (The Ad Spend)
Answer an ad-hoc question	Excellent	Yes, via Slack Q&A
Create a campaign from a brief	Yes (lands paused)	No — not its job
Remember last week's analysis	No — stateless	Yes — permanent record
Watch the account around the clock	No — acts only when prompted	Checks ~every 6 hours, 1,900+ detection algorithms
Prove which change caused a CPA spike	Narrative guess	Causal inference tied to the exact change
Log who changed what, when	No	Version-controlled history of every change
Blend Meta with Google, LinkedIn, TikTok, Reddit	No — single platform	Yes — cross-platform view
Approval before changes execute	No	Approve in app or Slack, everything logged

Why the gaps exist (and won't just get patched)

These aren't beta rough edges — they're properties of the protocol. MCP was designed to let an assistant reach tools during a conversation. Statelessness is the point: no server-side session, no accumulated record, no background process. That makes MCP safe and simple to adopt, and it's also why the assistant re-reads your account every session and answers "why" with narrative rather than evidence. Platforms compound the problem by exposing shallow histories of their own — see why ad platforms forget.

The Ad Spend is built to be the other half: a persistent layer that monitors Meta, Google Ads, LinkedIn, TikTok, and Reddit roughly every 6 hours with 1,900+ detection algorithms, keeps a permanent version-controlled record of every account change, traces performance moves to the exact change that caused them, and gates risky actions behind an approve-then-execute workflow in app or Slack. It complements the MCP rather than replacing it — the full comparison is at The Ad Spend vs. MCP.

FAQ

What can the Meta Ads MCP do?

It exposes 29 tools that let AI assistants pull performance reports and insights, create and edit campaigns, ad sets, and ads, and manage product catalogs on a connected Meta ad account. New objects it creates start paused; edits to existing objects apply immediately.

Does the Meta Ads MCP remember previous conversations?

No. MCP is stateless — each session starts from zero, with no memory of prior analyses, business context, or changes made. Any context you want considered has to be re-supplied every session.

Can the Meta Ads MCP monitor my account continuously?

No. It only executes when a human prompts it inside a chat session. It has no background process, no alerts, and no scheduled checks — continuous monitoring requires a persistent tool watching the account between sessions.

Is the Meta Ads MCP free?

Meta provides the server at no cost during the beta, but you need an AI client plan that supports remote MCP connectors, which generally means a paid Claude or ChatGPT tier.

Can it change my budgets without asking?

If write tools are enabled and you instruct it loosely, yes — edits to existing objects go live immediately with no draft stage. Restrict write tools in your client and keep a human approval step in front of changes.