Understanding Model Context Protocol: Connecting Your Software to AI (fusionauth.io)

The Model Context Protocol (MCP) formalizes how agent frameworks let large language models (LLMs) call out to external software — essentially a standardized JSON-RPC API and transport stack so tools work across different agent systems. MCP defines three actors (host/agent framework, client, server) and three layers (data/JSON‑RPC/transport). Two common transports are stdio (local process I/O) and HTTP streaming (remote), with tool calls like POST /mcp carrying JSON‑RPC "tools/call" requests and results often returned as Markdown or structured JSON. Crucially, MCP doesn’t change how LLMs reason — models remain stateless and the agent framework drives tool invocation by sending the full conversation to the model and interpreting its next chunk (user, agent, or tool response). Its significance is practical and strategic: MCP makes software automatable by any agent platform, so builders must treat agent access as a first‑class interface rather than an optional integration. That raises identity and security issues at scale — Prefactor argues MCP is fundamentally an identity problem. MCP Auth (for HTTP streaming) formalizes bearer tokens via OAuth with Dynamic Client Registration and PKCE: clients discover resource metadata, register dynamically, run the OAuth flow, obtain access tokens, and retry calls. The DCR requirement means not every OAuth server will interoperate out of the box. For developers, the takeaway is to design for standardized tool endpoints, robust auth flows, and careful permissioning to avoid creating the next wave of shadow IT.