LLMs don't know what's trending today, what the news said this morning, or why your app's rating dropped yesterday. Instead of writing a scraper per question, you can give an agent a few ready-made scrapers as tools through MCP (Model Context Protocol), and let it decide when to call them.

The setup

Apify runs a hosted MCP server at mcp.apify.com that exposes Actors (hosted scrapers) as tools. You choose which Actors it offers. For a "what's happening" assistant I add these (I maintain them):

  • fjordlabs/google-trending-now: today's trending searches per country
  • fjordlabs/google-news: headlines for a query, any country
  • fjordlabs/google-trends: interest over time for keywords
  • fjordlabs/google-play-reviews and fjordlabs/app-store-reviews: recent app reviews

In an MCP client config (Claude Desktop, Cursor, VS Code and others use this shape) it looks like this:

{ "mcpServers": { "apify": { "url": "https://mcp.apify.com/?actors=fjordlabs/google-trending-now,fjordlabs/google-news,fjordlabs/google-trends", "headers": {"Authorization": "Bearer <YOUR_APIFY_TOKEN>"} } } }

Check Apify's MCP docs for your client's exact format; the actors list is the part that matters.

What the agent can do now

Questions that used to need three tabs and a spreadsheet:

  • "What's trending in Germany and France today that isn't trending in the US?"
  • "Find this week's news about heat pump subsidies in the UK and summarise who's reporting it."
  • "Pull the last 7 days of 1–2 star reviews for com.spotify.music and group the complaints."

The agent picks the tool, fills in the input (countries, queries, star filter), and gets clean JSON back.

Why these work well as tools

  • Small, flat JSON. Each row is a few fields (term, rank, traffic, or headline, publisher, time), so results fit in the context window.
  • Sensible defaults. The agent can call them with one or two parameters.
  • Cheap and predictable. Pay per result ($0.001 per trending search, $0.0015 per article, $0.0001 per review), and failed rows are never charged, so a confused agent retrying doesn't cost much.

Without MCP

The same Actors work as plain function tools in any framework: call run-sync-get-dataset-items over HTTP and return the JSON to the model.

Disclosure: I built and maintain these scrapers. Questions and feature requests welcome in the comments. Links to Apify include my referral code.