Facebook Ad Library MCP

A small MCP server that scrapes the public Facebook Ad Library for competitive ad research — from your MCP client, with no Facebook account and no API token.

Install

git clone https://github.com/leyovco/facebook-ads-library-mcp.git
cd facebook-ads-library-mcp
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python -m playwright install chromium   # crawl4ai needs a browser

Register with your MCP client

Claude Code:

claude mcp add facebook-ads -- /abs/path/venv/bin/python /abs/path/facebook_ads_mcp_complete.py

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "facebook-ads": {
      "command": "/abs/path/venv/bin/python",
      "args": ["/abs/path/facebook_ads_mcp_complete.py"]
    }
  }
}

Restart the client.

Related MCP server: Facebook Ads Library MCP Server

Usage

"Search the Mexico Ad Library for 'automatización con inteligencia artificial',
 group by advertiser, and list the landing domains."

"Scrape https://www.facebook.com/ads/library/?active_status=active&ad_type=all&country=MX&q=nike&search_type=keyword_unordered
 with scroll_rounds=15 and summarise the creative angles."

Tools

Two tools:

• search_ad_library(query, country="MX", ...) — keyword search.

• scrape_ad_library_url(url, ...) — scrape any Ad Library URL you already have.

Both return structured records per ad:

Plus an advertisers histogram and the raw_markdown of the page so the model can pull anything the parser missed.

Configuration

• scroll_rounds (default 8)

• wait_seconds (default 8)

• country — ISO code (e.g., MX, US, ES)

• advertiser_page_id — target one Page's "all ads" view (less reliable than keyword search)

How it works

crawl4ai (AsyncWebCrawler) opens the Ad Library URL in headless Chromium, waits for the React app to hydrate, runs a scroll loop to trigger lazy-loaded cards, serialises the DOM to markdown, and a regex parser (_parse_ad_library_markdown) splits it on Library ID: boundaries and pulls the fields above.

Facebook answers the headless browser with HTTP 403 but still serves the rendered cards, so the tools judge success by whether cards parsed, not by status code.

License

MIT — original by Ramses Aguirre (RamsesAguirre777). Adapted and hosted under leyovco.