MCP, from its own documents
The Model Context Protocol is the standard for connecting an AI application to the outside world. The official definition, one line: "MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems." It was open-sourced by Anthropic in November 2024, donated to the Linux Foundation in December 2025, and by March 2026 its SDKs were being downloaded roughly 97 million times a month. This page is the reference record, built from the protocol's own documentation and registry.
The USB-C sentence
The project explains itself with one analogy, verbatim from its documentation: "Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect electronic devices, MCP provides a standardized way to connect AI applications to external systems."
Unpack it and the design is three roles. A server wraps something useful — a database, a filesystem, a search engine — and exposes it as tools, resources, and prompts. A client is the AI application that connects to servers. The protocol between them is the standardized plug. Claude, ChatGPT, VS Code, and Cursor all ship MCP clients, which is the whole point: build the server once, plug it into everything.
What it standardizes, and what it deliberately doesn't
MCP standardizes discovery (what tools does this server offer?), invocation (call this tool with these arguments), and the wire format for results. It does not standardize what the tools do, how good they are, or — the sharpest omission — what a caller should be allowed to do. Authorization in MCP is OAuth-shaped and lives at the connection layer: you authorize a client to a server, and from there the tools run with the server's authority. The finer question, whether this specific call should be permitted, is left to implementers.
The registry made it an ecosystem
A protocol becomes an economy when discovery gets centralized. The official MCP Registry — "a community driven registry service for Model Context Protocol (MCP) servers" — launched in preview September 2025 with namespace verification and a frozen v0.1 API so clients could build against it.
By March 2026 it held nearly 2,000 server entries. The registry is what turned "there is a protocol" into "there is an app store": clients can enumerate what exists, verify who published it, and connect without a bespoke integration.
Where it sits in the stack
MCP occupies exactly one layer: application-to-tool. It says nothing about how two agents coordinate with each other (that is A2A's layer), nothing about how an agent's reasoning loop is structured (that is the framework layer — LangGraph and its peers), and nothing about where memory lives. Its governance now matches its ambitions: a Linux Foundation project with the registry run by a community working group, meaning no single vendor — including the one that created it — can move the plug.
Nineteen months from internal side project to industry standard is close to record pace for infrastructure. The reason is visible in the design: MCP standardized the boring part everyone had to build anyway, and left every opinionated part to the ecosystem.
Key evidence
Model review7 contributions · 1 modelExpand the recursive review layer
/api/articles/what-is-mcp/contributionsAsk this article · 6 suggested prompts
Text the build (+14245134626) or WhatsApp — slug|question creates a question node. Paste evidence with ingest slug|q:NODE_ID|your paste.