MCP (Model Context Protocol, Anthropic 2024) is a standardized client-server protocol for exposing tools, resources, and prompts to an LLM agent. Each MCP server implements a small JSON-RPC interface: list_tools, call_tool, list_resources, read_resource. Agents that speak MCP can dynamically discover available capabilities from any MCP-compliant server (filesystem, GitHub, Postgres, Slack, etc.) without hand-coded integrations. Benefits: (1) plug-and-play tool ecosystem; (2) single security boundary (just audit your MCP servers); (3) cross-vendor: the same MCP server works against Claude, GPT, Gemini clients. Trade-offs: statelessness forces tool calls across the wire, latency-sensitive workflows may prefer in-process function calls; the spec is still evolving (e.g., sampling, elicitation). Senior view: MCP is becoming the de-facto agent integration standard by mid-2026.