What is MCP?
MCP stands for Model Context Protocol. It is an open protocol that gives AI applications one standard way to connect a language model to outside data and tools. An app such as a coding agent talks JSON-RPC to MCP servers, which expose tools, resources and prompts. Build a server once and every MCP app can use it.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
The host app (Claude Code, Cursor, a chat UI) creates one MCP client for each server it is configured to use.
The client sends JSON-RPC requests over stdio to a local process or over Streamable HTTP to a URL. Each request carries its protocol version and capabilities.
The server answers with what it offers: tools the model can call, resources the app can read, prompts the user can pick.
When the model wants to act, the client calls a tool, gets the result back, and the host puts it in front of the model and, where needed, in front of you.
Nothing is remembered between requests. State that must survive a call travels as an explicit id, so servers scale and restart freely.
MCP versus a plain API
An API is one service's own interface. To use ten services, an agent needs ten integrations, each with its own way to describe calls and return data. MCP puts one protocol in between. The server wraps the service and describes what it offers in a fixed shape; the app only has to speak MCP once. The specification compares this to the Language Server Protocol, which did the same for programming languages in editors: a language gets support everywhere by implementing one protocol instead of one plugin per editor.
The second difference is who the interface is for. A REST API is written for programmers. An MCP server is written for a model and the app around it: every tool carries a description and a JSON Schema the model reads before deciding to call it, and results come back as content the model can use directly.
Example: a request on the wire
Everything in MCP is a JSON-RPC 2.0 message. Here the client asks a server for the current weather:
{
"jsonrpc": "2.0",
"id": 2,
"method": "tools/call",
"params": {
"name": "get_weather",
"arguments": { "location": "New York" },
"_meta": {
"io.modelcontextprotocol/protocolVersion": "2026-07-28",
"io.modelcontextprotocol/clientCapabilities": {}
}
}
}
The _meta block is the part people miss. Since the 2026-07-28 revision, MCP is stateless: there is no opening handshake, and every request names the protocol version it speaks and the capabilities the client supports. The server can answer each message on its own, without remembering the one before. Earlier revisions opened with an initialize exchange, and clients that talk to both kinds of server detect which one they have reached.
What MCP covers
Three things flow from a server to the app: tools, which are functions the model can execute; resources, which are data the app attaches as context; and prompts, which are templates a person chooses. Clients can offer things back, such as elicitation, which lets a server ask you a question in the middle of a request. The specification also defines two transports, stdio for a server running as a child process and Streamable HTTP for one at a URL, and an authorization framework based on OAuth for the HTTP case. Optional extensions add more on top, including Tasks for long-running work and MCP Apps for interactive UI inside a conversation.
Common mistakes
- Treating MCP as an AI feature. It is a wiring standard. The model does not know MCP exists; it sees tools and results that the host app translated for it.
- Expecting a server to see the conversation. By design it gets only what the request contains. The full history stays in the host.
- Thinking tools run without a check. The spec says there should always be a human able to deny a tool call, and that a tool's own description of its behaviour is untrusted unless the server is.
FAQ
- What does MCP stand for?
- Model Context Protocol. The specification describes it as an open protocol for connecting LLM applications to external data sources and tools, using JSON-RPC 2.0 messages between hosts, clients and servers.
- Is MCP only for Claude?
- No. The specification is open, and any application can implement a host and any service a server. Claude Code, Cursor, Codex, Gemini CLI and VS Code all connect to the same MCP servers.
- What is MCP in AI, as opposed to other uses of the acronym?
- In AI, MCP means Model Context Protocol, the standard for connecting language model apps to tools and data. The letters have older meanings in other fields, which is why people add 'in AI' when they search for it.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.