MCP (Model Context Protocol)
The Model Context Protocol is an open standard that lets an AI assistant discover and call tools in other software at runtime. Published by Anthropic on November 25, 2024 and now maintained in the open.
MCP standardises the connection between an AI application and the systems it needs to act on. Anthropic’s own definition is “an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools”. Before it, connecting one tool to five AI products meant writing five integrations; after it, the tool describes itself once and any compliant client can read that description.
A call involves four parties, not two: the host application (Claude, ChatGPT, Cursor) contains an MCP client; the client calls a tool on an MCP server; the server reaches the underlying system with a credential you granted it; the answer comes back as data. The model never holds the credential — it sees a tool name, a description and a JSON schema for the arguments.
Versions are dates rather than semantic numbers, marking “the last date backwards incompatible changes were made”. There have been five: 2024-11-05, 2025-03-26 (which introduced streamable HTTP and made remote servers practical), 2025-06-18, 2025-11-25 and 2026-07-28, which removed the initialize handshake and protocol-level sessions and deprecated Roots, Sampling and Logging. A guide that does not mention the current revision is describing a protocol that no longer works that way.
MCP defines no approval step. Whether a change an agent proposes is staged for a human to click is a decision the individual server made, which is the single most important question to ask before connecting one to anything that spends money.