MCP — official introduction & architecture overview

The protocol’s own documentation (modelcontextprotocol.io), the primary source for what model-context-protocol actually specifies — previously the page had sources: [] after the financial-services source migrated out.

What it establishes

• Framing: MCP is “an open-source standard for connecting AI applications to external systems,” pitched as “a USB-C port for AI applications” — one standardized connector in place of bespoke per-integration glue.
• Participants (client-server):
  • MCP Host — the AI application (Claude Code, Claude Desktop, VS Code) that coordinates one or more clients.
  • MCP Client — a component that “maintains a connection to an MCP server and obtains context … for the MCP host to use”; the host spins up one client per server.
  • MCP Server — “a program that provides context to MCP clients,” local or remote.
• Two layers: a data layer built on JSON-RPC 2.0 (lifecycle, capability negotiation, the primitives) and a transport layer (connection, framing, auth).
• Server primitives: Tools (executable functions the model can invoke), Resources (contextual data sources), Prompts (reusable interaction templates). Each has */list discovery, retrieval, and (for tools) tools/call execution.
• Client primitives (the reverse direction): Sampling (server asks the host’s LLM for a completion — lets a server stay model-independent), Elicitation (server asks the user for input/confirmation), and Logging.
• Transports: stdio (local process, no network overhead) and Streamable HTTP (POST + optional Server-Sent Events for remote servers; OAuth recommended for auth).
• Stateful with capability negotiation: an initialize handshake (carrying a dated protocolVersion) negotiates which primitives each side supports before any work; servers can push notifications/*/list_changed so a client’s tool registry stays current.

Tier

T1 — the protocol’s own specification documentation (primary source).

Related

model-context-protocol · qmd · gbrain