Meta Ads via Claude: What the MCP Server Does and What You Can Build With It

What surprised me

I've been using MCP since Anthropic released it at the end of 2024. Initially mostly for filesystem, GitHub and a couple of databases. Nice, sped up our dev workflows, but nothing relevant for marketing.

A few weeks ago I installed the Meta Ads MCP server, an open-source implementation maintained by pipeboard.co. What surprised me wasn't the tool list itself. It was how quickly you go from "I ask Claude about my campaigns" to "Claude does that independently on a schedule". That's the threshold where tooling becomes operations.

Practical example from last week. I ask Claude: "Which ad sets in account X had a CPA above 45 EUR in the last seven days and are still active?" Claude calls get_adsets, then get_insights with the right date range, compares the CPA values, returns five ad sets. I follow up: "How was the trend over the last 30 days?" Claude re-fetches insights with a different range, builds a trend view. Then: "Pause the three with the worst trend." Claude asks me to confirm, I confirm, three update_adset calls, done. What would otherwise be 20 clicks in Ads Manager becomes three sentences.

MCP in 60 seconds

MCP stands for Model Context Protocol. It's an open JSON-RPC 2.0 protocol that defines how LLMs (Claude, ChatGPT, Cursor, VS Code) talk to external servers. These servers expose three things:

• Tools: functions the LLM can call (e.g. get_campaigns, pause_ad)
• Resources: data the LLM can read (e.g. schema definitions, configs)
• Prompts: reusable templates for complex interactions

Servers run either locally (via stdio pipes with the MCP host) or remotely (via Streamable HTTP). Local is the default for Claude Desktop. Remote is for SaaS tools.

What the Meta MCP server does

The pipeboard open-source server (github.com/pipeboard-co/meta-ads-mcp) currently exposes about 22 tools, grouped into five areas:

Account and hierarchy

get_ad_accounts, get_account_info, get_account_pages. You ask "which accounts do I have" and Claude lists them.

Campaigns, ad sets, ads

Get, get_details, create, update across all three levels. Important: create_campaign only accepts the new outcome-based objectives (OUTCOME_AWARENESS, OUTCOME_TRAFFIC, etc.). The legacy objectives are gone. You feel that when duplicating old setups, it doesn't work.

Creatives

upload_ad_image, create_ad_creative, update_ad_creative, get_ad_image. You can actually upload images and build new creatives from inside Claude. Including dynamic creative with multiple headlines and descriptions.

Audience and targeting

search_interests, get_interest_suggestions, validate_interests, search_behaviors, search_demographics, search_geo_locations. These are underrated. You tell Claude "find me interests for organic dog food buyers in their forties" and Claude tries a series of searches, validates reach, suggests a targeting list.

Insights and budget

get_insights is the most powerful tool. Performance metrics with breakdowns (age, gender, placement, device, etc.) and custom attribution windows. create_budget_schedule for time-bound budget pushes (Black Friday, season, launches).

What's missing in the current state: deeper custom audience management, no pixel or Conversion API setup, no full asset library handling. Read-heavy tasks are excellent, write-heavy works but is limited.

Setup: three paths

1. pipeboard.co (Remote OAuth). Visit the site, log in with Facebook, authorize ad account access, copy a connector URL into Claude Desktop. No local install. Fastest for non-techies.
2. Self-host locally. npm install or Python variant, your own Meta Developer App, System User Access Token with three scopes, edit Claude Desktop config. 20 to 30 minutes. Advantage: full control, token stays with you.
3. Markifact (commercial SaaS). Hosted version with prebuilt recipes and a UI layer. Monthly fee. Sensible if you manage multiple accounts and don't want to run anything yourself.

We use local self-hosted. Three reasons: token scope control, no external party with access to client accounts, and I can modify the server code when a tool is missing.

Claude Desktop config

If you go the local route, your claude_desktop_config.json looks roughly like this (path: ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

// claude_desktop_config.json
{
  "mcpServers": {
    "meta-ads": {
      "command": "npx",
      "args": ["-y", "@pipeboard/meta-ads-mcp"],
      "env": {
        "META_ACCESS_TOKEN": "EAAxxxxxxxxxxxxx",
        "META_AD_ACCOUNT_ID": "act_1234567890",
        "META_API_VERSION": "v22.0"
      }
    }
  }
}json

Restart Claude Desktop, in the chat you should see a small icon bottom-left indicating the server is connected. If not: tail -f ~/Library/Logs/Claude/mcp-server-meta-ads.log tells you why (usually token scope or wrong account prefix).

The first ten minutes

What I typically ask first, to test the connection and get the lay of the land:

• "Which ad accounts do I have access to?"
• "List the active campaigns in account act_xxx with current status and budget."
• "Which ad sets in campaign Y had a ROAS above 3 in the last 7 days?"
• "Compare performance of the top 5 ads in the last 14 days, grouped by placement."
• "Which of my ads had below 0.5% CTR in the last 3 days with more than 50 EUR spend?"

What you notice quickly: Claude understands domain language. "ROAS above 3", "CTR below 0.5%", "last 14 days" get translated into the right API calls without you needing to know field names. That's the actual magic. You speak your language, the LLM maps to API language.

Where it tips: combining MCP servers

Here's where it gets interesting. Claude Desktop can load multiple MCP servers at the same time. That means your setup doesn't just have Meta Ads access, it has every other MCP server you've connected.

Useful combinations:

• Meta Ads + Google Sheets MCP: export performance reports to a sheet
• Meta Ads + Telegram MCP: alerts and commands via Telegram bot
• Meta Ads + Slack MCP: daily standup in the marketing channel
• Meta Ads + GitHub MCP: derive creative briefs from PR descriptions
• Meta Ads + Filesystem MCP: match local CSV conversion data against Meta insights

As soon as you have two MCP servers, Claude can mediate between them. "Fetch performance data from Meta and write it to the sheet 'May 2026 Performance'" is a single command behind which 5 to 10 tool calls run across two servers.

Build idea: Telegram bot for daily ad updates

What we learned: formatting matters more than the data. A message with 30 numbers no one reads. A message with 3 numbers and a sentence "Heads up, campaign Y ate 40% more spend overnight" gets read. Let Claude interpret the data, not just fetch it.

Build idea: nightly setup audit

Build idea: anomaly alerts

Build idea: cross-platform reports

Risks and trapdoors

Token scope is delicate. If you give ads_management, the MCP server has full write rights on your account. Claude could theoretically pause every campaign if the prompt is badly worded. For read-only setups use only ads_read. For write access work with confirmation flows (the pipeboard MCP has this built in, custom implementations should have it too).

Auto-execution risk. If you tell Claude "clean up what's not performing" and Claude pauses twenty ad sets, three of which were in a learning phase, you have a problem. Confirmation before every write is mandatory, not optional.

Rate limits. The Meta Marketing API has strict tier-based limits. In aggressive setups (insights calls on 50+ campaigns every 15 min) you get throttled fast. Cache insights where possible, batch calls, respect headers. The pipeboard server does this reasonably well, but for custom setups you have to build it yourself.

Multi-account confusion. If you have access to multiple accounts, Claude can lose context. "Pause campaign X" is ambiguous if both accounts have a campaign X. We preface each prompt with the active account context ("We're working in account act_xxx") or restrict the MCP server per session to one account.

Currency and timezone. Meta delivers insights in the account currency and account timezone. When you ask Claude "last 7 days", the question is: from which timezone? Set it explicitly or document the assumption in the system prompt.

Data protection, the unsolved part. Meta data contains PII (email hashes in custom audiences, etc.). If Claude reads them and sends them to a third party (Anthropic API), that's a contractual issue with your clients. For enterprise setups we insist on Anthropic Workspace with DPA and EU region pinning. For smaller setups we clarify transparently in the DPA.

Which path for whom

• Marketers without setup appetite: pipeboard.co remote OAuth. 5 minutes setup, immediately ready.
• Agency with multiple client accounts: Self-hosted local, separated configurations per account, clear audit logs.
• In-house performance team: Self-hosted on a central server (e.g. a Cloud Run container) plus custom scheduling logic. The real build ideas above pay off here.
• Enterprise with compliance: Markifact or your own implementation with DPA. pipeboard.co can also work with an explicit contract.
• Curious, want to see first: Invest 30 minutes, connect pipeboard remote, talk to it for an hour, then decide.

Closing thought

We've been talking about "AI marketing" for 18 months, but most of what fell under that label was chatbots or auto-caption generators. MCP is the first technology that integrates what we actually do (analyzing campaigns, optimizing, reporting) into the LLM workflow. Not as a toy, but as an operations layer.

The interesting question is no longer "can AI take this over". It's: what do you do with the time you save once routine reporting is gone. Strategy. Creative briefings. Client conversations. The things humans are better at than any LLM stack.

If you haven't tried MCP yet, this is a good May afternoon. Go to pipeboard.co, connect, ask "what campaigns are running right now and what's the worst performer costing me". That first answer, in the tone you're used to asking questions in, is when it clicks.