Model Context Protocol (MCP): How Claude, Cursor, and AI Agents Connect to External Tools

Model Context Protocol (MCP), open-sourced by Anthropic in November 2024, is a transport-agnostic standard that lets AI agents discover and call external tools without custom integration code. An MCP server exposes three primitive types — tools (callable functions with JSON Schema input definitions), resources (read-only data sources the model can reference), and prompts (reusable prompt templates) — over JSON-RPC 2.0. The client connects via stdio (for local processes) or Server-Sent Events (for remote HTTP servers). When an agent calls tools/list, the server returns every available tool with its name, description, and input schema; the agent then issues tools/call with structured arguments to invoke them. This replaces the ad-hoc function-calling wiring that every LangGraph agent, CrewAI crew, and OpenAI Agents SDK runner previously required per-tool. By mid-2026, MCP hosts include Claude Desktop, Cursor, Cline, Zed, Windsurf, and production agents built on LangGraph and the OpenAI Agents SDK.

The key architectural implication is that tool graphs become inspectable without reading agent source code: the full tool list is discoverable at runtime via tools/list, and each tool call emits a span in OpenTelemetry-compatible traces with gen_ai.tool.name as the attribute. If you're building a multi-agent system where each agent connects to one or more MCP servers, the Agent Workflow Visualizer can scan your GitHub repository, detect agent-to-tool connections across LangGraph, CrewAI, AutoGen, and Google ADK, and render the relationship graph without any instrumentation changes.