Meta CLI and MCP for Claude Code: Skip the Developer App

Problem: For two years, running Meta ads from an AI agent meant picking a third-party MCP (Pipeboard, Ryze, Composio), creating your own Meta Developer App, and waiting four to six weeks for App Review. Most media buyers gave up and stayed in Ads Manager.

Quick Win: On April 29, 2026, Meta shipped the official meta ads mcp server and a Meta CLI that handle this for you. Standard Business OAuth, no Developer App, no review queue. Install the CLI globally and connect Claude Code in under fifteen minutes:

npm install -g @meta/ads-cli
meta auth login

This pillar post covers the practical operator workflow: when to use the CLI versus the MCP, the eight things you can now do from Claude Code that you could not last week, and how the new connectors slot into a real Shopify ad operation.

What Meta Just Shipped

The launch announcement (April 29, 2026) introduced two artifacts under the "Meta Ads AI connectors" umbrella:

1. Meta Ads MCP server, a hosted HTTP endpoint at https://mcp.meta.com/ads/<your-business-id>. Designed for desktop and chat-style clients (Claude Desktop, ChatGPT, any MCP-compliant agent) that wire MCP servers into a config file.
2. Meta CLI, a terminal binary installed via npm install -g @meta/ads-cli. This is the one Claude Code users actually run, because Claude Code already lives in your terminal.

Both connectors talk to the same underlying Marketing API surface, with full read and write access. You can query campaign performance, edit budgets, pause underperformers, manage product catalogs, and run signal diagnostics on Pixel and CAPI.

The headline change versus the previous two years: no Meta Developer App approval required. Launch coverage clocks setup at five-to-seven minutes for the MCP and ten-to-fifteen for the CLI, versus the old "twenty-five-plus minutes and a multi-day App Review" path. Meta's own announcement frames it as "no developer credentials, API setup, or coding required."

The connectors are free during the open beta. Long-term pricing is unannounced as of May 2026. New to MCP itself? Start with our MCP fundamentals guide and come back here.

Meta CLI vs Meta MCP: Which One You Actually Need

This is the central confusion in every SERP result we found, so we are settling it first.

Use the Meta CLI when you are inside Claude Code, in the terminal, and you want Claude to run ad operations as part of a coding or analysis session. The CLI exposes commands like meta campaigns list, meta insights run, and meta budgets set that Claude can call as shell commands. No JSON config, no restart cycle. This is the right choice for ninety percent of operator workflows.

Use the Meta MCP server when you are in Claude Desktop, ChatGPT, or another non-terminal MCP client, or when you want a long-lived natural-language interface that does not require Claude to spawn shell processes. The MCP runs as an HTTP transport. Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "meta-ads": {
      "url": "https://mcp.meta.com/ads/<your-business-id>"
    }
  }
}

Restart Claude Desktop and the connector authenticates through Meta Business OAuth in your browser. Both surfaces speak the same API, so capabilities are identical. The difference is purely about where the agent lives.

For mixed-mode operators (terminal during the week, Desktop on weekends for quick checks), running both simultaneously is fine. They do not conflict. If you run more than two or three MCPs at once, enable MCP Tool Search so Claude only loads the connector when you actually need it. Without tool search, three MCPs can burn around twelve thousand context tokens at session start. With it, the same three drop to roughly six hundred until invoked.

Installing the Meta CLI in Claude Code

The Meta CLI Claude Code workflow is the path most operators will run. The CLI ships as an npm package. Node 18 or higher is required.

# Install globally
npm install -g @meta/ads-cli
 
# Authenticate via Meta Business OAuth (opens browser)
meta auth login
 
# Confirm the connection and list your business accounts
meta accounts list

The OAuth flow requests standard ad-account scopes: ads_management, ads_read, business_management. If your account is part of a Meta Business Manager with multiple ad accounts, meta accounts list returns each one with its act_<id> identifier. You can pin a default for your shell session:

export META_ACT_ID=act_1234567890

From this point, Claude Code can run any meta command directly. Tell Claude what you want and let it pick the command:

"Pull the last seven days of ROAS by ad set for our prospecting campaign and flag any with frequency over 3.0."

Claude reasons about the request, calls meta insights run with the right time range and breakdown, parses the response, and writes a summary. There is no MCP wiring step, because Claude Code treats the CLI like any other command in your shell. If you have permissioned meta commands at the project level, Claude does not need to ask before each call. See our permission management guide for the exact pattern.

Setting Up the Meta Ads MCP Server

For Claude Desktop or ChatGPT users, the official Meta Ads MCP server is the right path. The endpoint is https://mcp.meta.com/ads/<business-id>. You retrieve your business ID from Meta Business Suite under Settings → Business Info, or via the CLI: meta accounts list --json | jq '.[0].business_id'.

For Claude Desktop on macOS, edit ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows, the file is at %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "meta-ads": {
      "url": "https://mcp.meta.com/ads/123456789012345"
    }
  }
}

Restart Claude Desktop. The first natural-language ad query triggers a browser-based OAuth handshake. Tokens are issued and refreshed by Meta's auth service. Per the official help center, no local token storage is required, and you do not need to maintain a long-lived refresh token in a .env file.

This is the meaningful security improvement over the unofficial Pipeboard or Composio paths, where you either store a Developer App secret locally or trust a third-party service to broker auth. Meta's MCP is a first-party endpoint, so the auth boundary is the same one you already trust for the Ads Manager web UI.

Eight Things You Can Now Do From Your Terminal

The MCP and CLI both expose the full Meta Marketing API in Claude Code, a broader surface than most third-party MCPs covered. Here is the operator-grade list that matters for a real Shopify or DTC media operation:

1. Pull insights at any breakdown. ROAS by ad set, frequency by ad, CPM by placement, CTR by creative. Claude turns natural language into the right Insights API call. Ask "show me yesterday's performance grouped by audience" and the right breakdowns=age,gender,country parameters appear without you knowing the API surface.

2. Edit budgets in batch. "Cut budget on every ad set with a CPA over $80 in the last three days by twenty percent" runs as a single conversation. Claude pulls insights, filters, calls meta adsets update for each match, and reports back. This used to require a custom Python script or a manual loop in Ads Manager.

3. Pause underperformers. Same shape as budget edits. The cost of a bad ad running another twelve hours overnight, while you sleep, was the most common pain point cited in pre-launch Reddit and X threads about Meta MCP demand. The official connector solves it.

4. Launch lookalike audiences. Audience Manager via API. Claude can read a list of your top-spending customers from a Shopify export, push it as a custom audience, and request a 1% lookalike in one prompt.

5. Audit Pixel and CAPI signals. Signal diagnostics is one of the four headline capabilities Meta called out. Claude can run meta signals diagnose against an event source group, surface deduplication problems between Pixel and CAPI, and flag missing event_id parameters. This is the post-iOS-14 work that consultants charge $3K-$8K to do once.

6. Manage product catalogs. The Shopify-Meta sync surface. If your product catalog has stale prices or broken image URLs, Claude can read the catalog feed errors, group them by product type, and tell you which ones are blocking spend. Combine this with Shopify ad-side fixes for a full catalog-health pass in one session.

7. Batch creative tests. Upload twenty new ad variants, attach them to the same audience, set a $20/day learning budget per variant, and let the data accumulate. Claude scripts this as a single workflow. No spreadsheet, no QA from a junior media buyer at 2am.

8. Generate weekly reports automatically. Combined with Claude Code's hooks, a Sunday-morning slash command can fire meta insights run across every active campaign, summarize movers, flag anomalies, and write a Markdown report. Same job a $500/month reporting tool does, running in your own terminal for free.

The point of the list is not the eight bullet points. The point is that the Meta Ads MCP turns operating ads into a tool-using session, not a UI-clicking session. The mental model shift matters more than any single capability.

Connecting Meta to Your Shopify Stack

Running Meta ads in isolation hits a ceiling fast. The real leverage is the loop: a creative tool generates the assets, Meta distributes them, Shopify converts the traffic, and Meta's CAPI feeds clean conversion data back into the optimization model. Claude Code can now sit in the middle of that loop and operate every layer.

The two adjacent pillars in this stack:

• Shopify side: Shopify shipped its official AI Toolkit on April 9, 2026, three weeks before Meta. The toolkit's Dev MCP and Storefront MCP let Claude edit product detail pages, fix broken metafield definitions, audit checkout extensibility, and pull live analytics. When a Meta campaign tanks, the diagnostic question "is it the ad or the landing page" can be answered in the same session.
• Creative side: Higgsfield's video and image MCP, launched April 30, gives Claude a programmable creative surface. Generate twelve product video variants, push them to Meta as a creative test, and watch fatigue curves play out three days later from the same terminal session.

This is the AI-Powered Shopify Operations stack: Higgsfield generates, Meta distributes, Shopify converts, and Claude Code is the operator running all three.

The complication is that running Meta ads against a real Shopify catalog is not a tooling problem at this point, it is a playbook problem. Andromeda creative cadence, post-iOS measurement frameworks, diagnostic trees for when CPM doubles overnight, the right cross-channel attribution model for a sub-$5M store. The Shopify Kit ($199) ships eleven paid-ads playbooks built for exactly this stack, alongside thirty-one SEO files, ten CRO files, and eight retention flows. If you are wiring Meta MCP into a real store, that is the operator framework that closes the loop.

Pricing, Limits, and the Beta Caveats

The connectors are free during the open beta. Meta has not committed to long-term pricing. Reasonable bet: connector usage rolls into existing Marketing API rate limits without a per-seat charge, because the connector is a thin layer over the API surface that big advertisers already access.

The hard limits to know:

• Marketing API rate limits apply. The connector does not bypass them. Standard accounts get an hourly call budget that scales with ad spend. For most sub-$50K-per-month accounts, the limit is not a practical constraint. Above that, batch your Insights calls.
• No autonomous-agent layer. The connector does not "run your account for you" in any persistent sense. Claude executes commands you ask for. There is no scheduled-task surface in the official connector itself. Use Claude Code hooks or your own cron.
• No Business Manager auto-discovery for client agencies. Each client account requires its own OAuth handshake. Per-client token isolation is light. Agencies operating many accounts under one Claude Code session need to swap auth contexts manually.
• Beta capability gaps. As of May 2026, advanced placements like Audience Network reservation buys, branded content tools, and some lift studies are not exposed. Standard auction campaigns, conversion campaigns, and catalog ads cover the operator workflow Meta demoed at launch.

Why This Replaces Pipeboard, Ryze, and the Other Unofficial MCPs

For two years, the practical answer to "how do I run Meta ads from Claude" was Pipeboard's meta-ads-mcp. It is a solid open-source project, MIT-licensed, with twenty-nine MCP tools and an 848-star repository. Composio and Ryze each shipped their own variants. Improvado built one as part of a broader BI integration. They worked. They also required, in every case, that you create your own Meta Developer App, navigate App Review (multi-day to multi-week, depending on permission scope), and manage your own access tokens.

Meta's official launch obsoletes that path for most operators. Three reasons:

1. Auth surface is first-party. Token issuance, refresh, and revocation are owned by Meta. The previous third-party path had you trusting either an open-source repo or a SaaS broker to hold a token that could read and write your ad spend.
2. No App Review. The lift between "I want to try this" and "I am running campaigns from Claude" drops from days to minutes.
3. Capability ceiling is the full Marketing API. Pipeboard exposes a curated twenty-nine tools. Composio exposes a different curated set. The official connector exposes the API surface itself, which means future Marketing API features show up in the connector immediately, not on a community-MCP roadmap.

When does an unofficial server still make sense? Three scenarios. If Meta Ads is one of twenty toolkits in the same agent, Composio's Tool Router beats running ten separate official MCPs. If you need autonomy or scheduling beyond Claude Code's hook system, a self-hosted MCP gives you a process you control. If compliance requires per-client token isolation that Meta's connector does not yet expose, a self-hosted server with explicit token boundaries wins. Otherwise, default to official.

Your First Hour With Meta CLI

Concrete walk-through. Pour coffee, open a terminal in a project directory, and run:

# Install
npm install -g @meta/ads-cli
 
# Authenticate
meta auth login
 
# Confirm
meta accounts list

Pin your account. In Claude Code, ask:

"Pull the last seven days of spend, ROAS, CPA, and frequency for every active campaign on act_<id>. Format as a Markdown table sorted by spend. Flag anything with frequency over 3.0 or CPA more than 30% above the campaign average."

Claude calls the right insights endpoint, parses the response, computes the deltas, and writes the table. Fifteen seconds. The table tells you which campaigns to look at next.

For each flagged campaign, the follow-up:

"Pull the last fourteen days of insights at the ad-set level for campaign <id>. Identify which ad set is dragging the average. If frequency is over 4.0, suggest pausing the worst-performing ad inside it."

Claude pulls the data, flags the ad set, names the ad. You confirm. Claude pauses it via meta ads update --status PAUSED <ad-id>. Total elapsed time from logging in to the first pause: roughly twenty minutes if you have never seen the CLI before.

The hour ends with a budget reallocation. Tell Claude to take ten percent of the budget from the worst-performing ad set in the campaign and add it to the best, capped at a fifteen percent total daily change so the algorithm does not lose its learning phase. Claude does the math, runs the two meta adsets update calls, and writes the change log to a Markdown file in your repo.

That report is the artifact of the session. Commit it. Three weeks later, you have a paper trail of every adjustment, what data drove it, and what the result was. The Ads Manager UI gives you none of that.

The shift is not "Claude clicks Ads Manager faster." It is that Meta ad operations becomes a versioned, scripted, repeatable practice. That is what Pipeboard and Composio always promised but rarely delivered, because auth friction blocked adoption. Meta's own connector removes that friction.

Where to Go Next

The Meta CLI and Meta MCP are the distribution layer of an operator stack that finally has an official version of every component. Wire them up next to the rest:

• Set up the Shopify AI Toolkit so Claude can edit your store, validate Liquid templates, and pull live analytics in the same session it adjusts your Meta budgets.
• Add Higgsfield MCP for programmable creative generation. Twelve variants in a Sunday session, tested in Meta on Monday, performance read by Wednesday.
• Browse our curated MCP list for the auxiliary connectors (Klaviyo, Triple Whale, GA4) that round out the operator stack.
• If you operate other social channels alongside Meta, our social media MCPs guide maps the rest of the surface.

The Shopify Kit ($199) ships the playbooks behind this stack: 11 paid-ads files for Andromeda creative cadence and post-iOS measurement, 31 SEO files, 10 CRO files for the landing-page side of every Meta campaign, and 8 named playbooks including the cross-channel diagnostic trees you run when a campaign tanks. Meta's official MCP gives you the hands. The Shopify Kit gives you the playbook. Together, that replaces the senior media buyer plus CRO consultant plus analytics retainer most $500K-$10M Shopify stores pay $5K-$15K a month to assemble piecewise.