MCP

MCP, or the Model Context Protocol, is an open protocol — released by Anthropic on 25 November 2024 — that lets an AI agent discover and call a server's tools at runtime rather than through integration code written in advance. It runs over JSON-RPC 2.0: a tools/list request returns the tools the caller may invoke, each carrying a name, a description, and an inputSchema that must "be a valid JSON Schema object" (2020-12 by default), plus an optional outputSchema for the result; tools/call then invokes one. Servers are reached over stdio, where "the client launches the MCP server as a subprocess", or over Streamable HTTP, a single POST endpoint; HTTP servers that support authorization act as OAuth 2.1 resource servers, and clients "MUST implement PKCE". Failures split two ways — JSON-RPC errors for malformed requests, and tool execution errors returned with isError: true, giving the model "actionable feedback" to self-correct against. As of the 2026-07-28 revision there is no initialize handshake — "There is no negotiation handshake"; each request declares its own protocol version, and "MCP has no protocol-level session". The protocol itself "has no concept of a state handle"; as non-normative guidance, a server needing state across calls should return an explicit handle and accept it as an argument on later calls. Compare EPP, the protocol registrars use with registries, and x402 for agent payments; Why MCP Exists When We Already Have REST APIs and CLIs covers the design rationale.