Meta Ads MCP and CLI: Manage Facebook & Instagram Campaigns From Claude or ChatGPT

Meta released two official surfaces for AI-driven ad management on April 29, 2026 — both announced simultaneously on the Meta for Business newsroom and the Meta for Developers blog.

1. The Meta Ads MCP server

A hosted Model Context Protocol endpoint at https://mcp.facebook.com/ads. You add it to Claude Desktop or ChatGPT as a connector, sign in with your Meta Business account, and the chat window now has authorized access to your ad accounts. No Developer App registration. No long-lived access tokens to babysit.

2. The meta-ads-cli

A local Python CLI (Python 3.12+, install via pip or uv) that exposes the same 29 Marketing API tools as scriptable commands. It is the path Meta recommends for Claude Code, CI runners, and shell-based workflows.

Both surfaces wrap the Meta Marketing API. Both run through Meta Business OAuth. Both create resources in PAUSED status by default — a human still has to flip them live. And both are free during the open beta.

Why this announcement matters

Until April 29, connecting an AI agent to a Meta ad account meant registering a Developer App, surviving app review, managing system user tokens, and praying that none of your automated edits looked enough like bot traffic to trigger a ban. Bryan Cano's reaction tweet captures why marketers cared immediately.

Both the MCP server and the CLI expose the same 29 tools. They cluster into five buckets — useful to know before you write a prompt or a script.

Campaign management (5 tools)

List, create, update, duplicate, and delete campaigns, ad sets, and ads. This is the cluster that produces new objects in your account — and the one that always defaults to PAUSED.

Product catalogs (10 tools)

The biggest single bucket. Build and manage product catalogs and product sets, push feed updates, and read catalog health. This is what makes the connectors interesting for ecommerce: a single agent can audit a catalog, fix missing fields, and push corrected data without a CSV round-trip.

Accounts (3 tools)

List ad accounts, switch the active account, and inspect account-level metadata.

Dataset quality (4 tools)

Pixel firing status, conversion API health, signal diagnostics. The cluster that catches "why is my CPA suddenly broken" before you waste a budget cycle.

Insights (7 tools)

The reporting cluster: spend, impressions, CTR, ROAS, frequency, breakdowns, and custom date ranges. This is where most of the daily-driver work happens — and it is the cluster the Ads Uploader team has publicly called "well-served" by AI, as opposed to launch workloads which are still "structurally fragile."

There are three install paths depending on where you want the agent to live.

Path 1 — Claude Desktop

1. Open Claude → Customize → Connectors
2. Click + Add custom connector
3. Paste the endpoint: https://mcp.facebook.com/ads
4. Sign in through Meta's official OAuth screen
5. Pick a scope tier (read-only / read-write / financial)
6. Toggle the connector on inside any conversation

Path 2 — ChatGPT (paid tiers)

1. Settings → Apps → Advanced settings
2. Enable developer mode
3. Add the Meta MCP URL https://mcp.facebook.com/ads
4. In a chat, switch to Developer Mode and select the Meta Ads connector
5. Approve through Meta's auth page

Path 3 — CLI (Claude Code, CI, shell scripts)

# Requires Python 3.12+
pip install meta-ads-cli

meta auth login
meta config set default_account act_1234567890

# First sanity check
meta ads campaign list

The CLI reads tokens and account IDs from environment variables, so you can keep secrets out of shell history and out of any committed config. Combined with meta auth login, this gives Claude Code the same authority as the chat connectors — but inside a repo, alongside your code.

Quick smoke test prompt

Once connected, the fastest way to know it works:

"Show me my top 5 campaigns from the last 7 days by ROAS, with spend and conversions."

If you get a structured answer with real numbers from your account, the OAuth handshake worked.

Within the first 48 hours, three workflow patterns emerged across Twitter, Reddit, and the early write-ups from agencies and tooling companies.

1. Daily anomaly watchdogs

A scheduled CLI job (or a daily Claude run) that compares today's spend, CPM, and conversion volume against a 14-day baseline. Anything more than two standard deviations out gets summarized into a Slack message. The DataAlly marketer's guide walks through the pattern with concrete prompts.

2. Catalog and Pixel health audits

This is the use case that only works because of the catalog (10 tools) and dataset-quality (4 tools) clusters. Agents scan for missing product fields, broken Pixel events, and unfiring CAPI endpoints — work that previously meant a human clicking through Events Manager.

3. Bulk paused-campaign creation from briefs

Marketers feed the agent a campaign brief (objective, audience, budget, creative URL) and ask it to draft N variants. Because everything ships PAUSED, this is safe to run before review. Common Thread Collective's analysis frames this as "reporting workloads are well-served and launching workloads are structurally fragile" — which is a fair read of where the connectors are strongest today.

What is not working yet

• Custom conversions and Cost-per-Add-to-Cart are missing from the insights cluster
• Coverage is shallower than the full Marketing API — power users still hit endpoints the connectors don't expose
• Rollout is wave-based — not every account sees the connector option immediately
• Meta-only — there is no equivalent for Google, TikTok, LinkedIn, or X yet

The biggest historical risk with AI-driven ad automation has been account suspension for unsanctioned API usage. The official connectors materially change that risk profile, but they don't eliminate it.

Three OAuth scope tiers

• Read-only — query campaigns, insights, catalogs, and dataset quality
• Read/write — also create, edit, duplicate, and pause objects
• Financial — also adjust budgets and bid strategies

Give the agent the smallest scope that does the job. Reporting agents should never get the financial scope.

What lowers risk on the official path

• Meta's own OAuth flow — the consent page is unmistakably Meta's
• Traffic flows through official endpoints — not flagged as non-human
• PAUSED-by-default on every created object — humans flip the switch
• No personal access tokens floating in .env files for someone to leak

What is still risky

• Rate-limit thrash — an agent in a tight loop still gets throttled
• Constant automated edits — high-frequency budget churn still looks suspicious
• Mixing official and unofficial connectors — third-party MCP servers using personal tokens are still in the same risk zone they were in before

A practical safety pattern

1. Start every new agent on the read-only scope for at least a week
2. Promote to read/write once you trust the tool selection
3. Promote to financial only for narrow, well-tested workflows
4. Keep a staging campaign with a tiny daily budget that the agent practices on before any real account work

Yes — install the read-only scope today, even if you're not ready to let an agent spend money. The reporting cluster alone replaces a meaningful chunk of weekly CSV-export-and-pivot work, and the OAuth path is the lowest-risk way to give Claude or ChatGPT access to your account.

The launch surfaces are unmistakably v1: 29 tools is a lot, but it's still a subset of what the Marketing API can do. Custom conversions are missing. Wave rollout means some accounts wait. Launch workloads are still better as drafts than as live decisions.

But the architectural shift is real. "Manage my Meta Ads account by talking to Claude" is no longer a hack involving a Developer App and a system user token — it's a connector toggle. That's the kind of change that compounds.

Primary launch sources

• Meta for Business — Introducing Meta Ads AI Connectors — Official announcement, April 29, 2026
• Meta for Developers — Introducing Ads CLI — CLI launch post and command reference
• Manage Ads from an AI Agent — Meta Business help center setup guide
• Meta Ads — MCP Server listing — Tool catalog and scope reference

Community reactions and analysis

• Bryan Cano's launch-day tweet — The viral marketer's reaction quoted above
• Meta Ads MCP and CLI: Inside Meta's Official AI Connectors — mcp.directory's deep dive
• Meta Just Launched AI Connectors for Ads — Common Thread Collective's ecommerce analysis
• Meta Ads MCP and CLI: How to Connect Meta to Claude or ChatGPT — DataAlly's marketer setup walkthrough
• Meta Ads MCP + CLI: Skip the Developer App — Claude Fast's developer-focused walkthrough

Related context

• pipeboard-co/meta-ads-mcp on GitHub — The leading community-built Meta Ads MCP server (predates the official launch)