MCP Server

An Apify MCP (Model Context Protocol) Server is a specialized actor that exposes Apify actor functionality as callable tools for AI agents, large language models, and AI-powered development environments like Claude, ChatGPT, Cursor, Windsurf, and VS Code Copilot. MCP is an open standard (developed by Anthropic) that defines how AI systems discover and invoke external tools, and Apify MCP servers implement this protocol to bridge the gap between AI agents and real-world data extraction, web scraping, and automation capabilities. MCP servers matter because the AI agent ecosystem is rapidly growing beyond simple text generation into tool-using agents that interact with external services. When an AI agent needs to scrape a website, extract data, monitor prices, gather intelligence, or automate a workflow, it needs a standardized way to discover available tools, understand their inputs, invoke them, and process their outputs. MCP provides that standard, and Apify MCP servers provide the tools. There are over 80 MCP intelligence servers on the Apify Store, covering domains from financial crime screening to cybersecurity intelligence to competitive analysis. An MCP server runs as an Express HTTP server with StreamableHTTPServerTransport, accepting tool calls via POST /mcp (or alternatively via SSE at /sse for older clients). Each MCP bundles 5-18 related actors into themed intelligence tools. For example, a financial crime MCP might expose tools like screen-entity, check-sanctions, analyze-transactions, and assess-risk. When an AI agent calls one of these tools, the MCP server runs the underlying Apify actors via Actor.call(), aggregates results with domain-specific scoring algorithms, and returns structured intelligence in a format the AI agent can understand and present to users. To configure an MCP server for use with Claude Desktop, add it to your claude_desktop_config.json: { 'mcpServers': { 'apify-intelligence': { 'url': 'https://your-actor.apify.actor/mcp' } } }. For Cursor, add it to .cursor/mcp.json. For programmatic access, use the MCP client SDK: import { Client } from '@modelcontextprotocol/sdk/client'; const client = new Client({ name: 'my-app' }); const result = await client.callTool({ name: 'screen-entity', arguments: { entityName: 'John Doe' } }); To build an MCP server actor: create an Express app, attach the MCP transport, register your tools with their input schemas and handler functions, and deploy with Standby Mode enabled so the server stays running persistently. Each tool handler typically validates the input, calls one or more Apify actors with Actor.call(), processes the results, and returns structured data. The key design decision is which actors to bundle together — group actors by domain or use case so that the MCP provides a coherent set of related capabilities. Common mistakes when building MCP servers include not enabling Standby Mode, which means the server has a 2-10 second cold start on every request — unacceptable latency for interactive AI agent usage. Another mistake is exposing too many tools in a single MCP (more than 20), which overwhelms the AI agent's tool selection and leads to poor tool choice accuracy. Keep each MCP focused on a single domain with 5-18 tools. Not implementing proper error handling is also problematic: if an underlying actor fails, the MCP should return a clear error message rather than crashing, so the AI agent can retry or try an alternative approach. MCP servers represent the next phase of the Apify ecosystem — transforming data extraction actors from standalone tools into AI-native capabilities that any intelligent agent can use. Related concepts: Standby Mode, Actor, Actor Run, Webhook, PPE.