What is MCP (Model Context Protocol)?
MCP (Model Context Protocol) is an open standard that gives AI models a uniform way to use external capabilities: tools they can call, resources they can read, and prompts they can reuse. A client such as Claude, ChatGPT, Cursor, or VS Code connects to any MCP server and discovers what it offers at runtime — so one integration works across every MCP-capable app, instead of being rebuilt for each one.
The problem MCP solves
Before MCP, connecting an AI product to a service meant a bespoke integration for that one pairing — a ChatGPT plugin, a Claude-specific tool, a custom agent harness. With M assistants and N services, the industry was writing M × N integrations, each with its own auth story and each obsolete the moment either side changed.
MCP collapses that to M + N. A model-side application implements the protocol once as a client; a service implements it once as a server; every client then works with every server. It is the same shape as USB or LSP: standardize the connector, and the ecosystem composes.
How MCP works
MCP is a client–server protocol over JSON-RPC. Local servers run as a child process and speak stdio; remote servers live at a plain HTTPS URL and speak the streamable HTTP transport. When a client connects, it asks the server what it offers and gets back typed, self-described capabilities the model can read.
Those capabilities come in three primitives:
- Tools — functions the model can call, each with a JSON Schema for its arguments and a description written for the model to reason about.
- Resources — data the client can read into context: files, documents, records.
- Prompts — reusable, parameterized prompt templates the server ships alongside its tools.
Who supports MCP
Anthropic introduced MCP in November 2024 and open-sourced the specification. It has since become the industry default: Claude (Desktop, Code, and web), ChatGPT custom connectors, the Gemini CLI, Cursor, VS Code, Windsurf, and Codex all speak it, and community directories track thousands of public servers.
Because the protocol is open, none of that adoption locks you in. A server you ship today is reachable from every one of those clients — and from the ones that have not shipped yet.
Common questions
Who created MCP?
Anthropic introduced MCP in November 2024 and open-sourced the specification. It is now an open standard with client support from OpenAI, Google, Microsoft, and most AI-native editors, plus thousands of community-built servers.
Is MCP only for Claude?
No. Any client that implements the protocol can talk to any MCP server — ChatGPT custom connectors, Cursor, VS Code, Windsurf, and the Gemini CLI all connect to the same server without server-side changes.
How is MCP different from function calling?
Function calling is a per-application feature: each app defines its own tools and wiring, and none of it transfers. MCP standardizes how tools are discovered, described, and invoked, so one server is portable across every MCP client instead of being reimplemented inside each application.
Do MCP servers have to run locally?
No. Local servers speak stdio, but remote servers speak streamable HTTP at an ordinary URL — which is how hosted platforms serve them, with authentication, rate limits, and observability handled server-side instead of on every laptop.