Apify has established a dedicated Model Context Protocol (MCP) server at https://mcp.apify.com, enabling AI clients like Claude Desktop and Cursor to execute web scraping actors directly through plain-English prompts. This integration bypasses traditional API complexity by allowing the LLM to select and run specific actors, such as contact extractors or sitemap parsers, without requiring the developer to write custom orchestration code for each task.

The Configuration Strategy

The core of this setup involves pointing the MCP client to the remote server while using query parameters to restrict available tools. By appending ?tools= to the URL, developers can limit the agent’s scope to specific actors, such as pennydata/website-contact-extractor or pennydata/sitemap-extractor-plus. This constraint is critical for maintaining predictability; without it, the agent defaults to Apify’s entire store, potentially selecting inappropriate or expensive tools for simple tasks. For Cursor users, the configuration can be injected via the settings menu or automated using the Apify CLI command apify mcp install cursor. Claude Desktop users can add the remote URL as a connector, though those on builds that only support local servers must use an npx fallback with the @apify/actors-mcp-server package. Authentication is handled via OAuth for the remote server, eliminating the need to store API tokens in configuration files, though a Bearer token header remains an option for local setups.

Real-World Execution Examples

The article demonstrates two practical use cases: extracting contact information and parsing sitemaps. In the first example, the agent runs the website-contact-extractor on domains like apify.com and mozilla.org, returning a structured table of contact pages, forms, and emails. The actor charges $0.002 per site where data is found, with failed or empty results incurring no cost. In the second example, the sitemap-extractor-plus processes docs.apify.com, grouping URLs by path segment and identifying lastmod dates, priced at $0.0004 per unique URL.

Key Takeaways

  • Use the ?tools= parameter to whitelist specific Apify actors, preventing the LLM from hallucinating or selecting irrelevant scraping tools.
  • OAuth is the preferred authentication method for the remote MCP server, keeping API tokens out of local config files.
  • Agents do not verify data integrity; developers must instruct them to flag anomalies, such as joke emails or missing lastmod dates in sitemaps.
  • Spend control is native to the actors, with parameters like maxUrls acting as hard caps on API costs during agent execution.

The Bottom Line

While the MCP integration lowers the barrier for non-developers to leverage structured web data, the lack of agent verification for data integrity means human oversight remains essential. Developers should treat these agents as powerful but naive executors, strictly bounding costs and validating outputs against the raw actor data rather than trusting the LLM's summary.