MCP

MCP defines a client-server architecture. The AI application (Claude Desktop, Cursor, a custom agent) is the client. Tool providers (databases, APIs, file systems, dev tools) are servers. MCP lets any client talk to any server without custom integration code. As of 2026, the MCP registry lists 4,000+ servers covering GitHub, Slack, filesystem, web search, Kubernetes, every major SaaS, and more.

MCP uses JSON-RPC over stdio, SSE, or WebSocket. Servers expose three primitives: resources (readable state), tools (callable functions), and prompts (predefined templates). Clients connect to servers at startup, list capabilities, and surface them to the model. The model decides when to invoke tools; the client executes and returns results. MCP solves the N×M integration problem: N AI applications × M tools collapse to N + M MCP implementations. The protocol is spec-first and open; Anthropic published the SDK in multiple languages. Cursor, Cline, Claude Desktop, Zed, and most AI coding agents now ship MCP support.

MCP specification v1 defines the message shape (JSON-RPC 2.0 envelope), the capability negotiation handshake, and authorization patterns. Servers run as separate processes for isolation (stdio) or as remote services (HTTP+SSE or WebSocket). The registry at registry.modelcontextprotocol.io lists 4,300+ servers. Security concerns: prompt injection through tool descriptions, unintended tool invocation, cascading data exfiltration through chained tool calls. Mitigations: explicit user approval for each tool invocation (Claude Desktop's confirmation dialog), allowlists, sandboxed execution environments. MCP is gaining competitive alternatives (OpenAI Functions, LangChain Tools) but its open-protocol status makes it the interoperability winner.