What does MCP stand for?

Model Context Protocol. It is an open protocol for connecting AI models and agents to external tools and data sources.

Is an MCP server a web server?

Not necessarily. Most local MCP servers are command-line programs that the agent starts and talks to over stdin/stdout (the stdio transport). Remote MCP servers use HTTP.

Which apps support MCP?

Claude Code, Claude Desktop, Cursor, OpenAI Codex, VS Code, Windsurf, and many other agents and IDEs support MCP servers.

How can I see what an MCP server is doing?

Put a recorder like RelayOrb in front of it. RelayOrb relays the traffic unchanged and saves every request and response so you can inspect, replay, or test it later.

What is MCP, and what is an MCP server?

Updated 2026-10-05

MCP, the Model Context Protocol, is an open standard for connecting AI agents to tools and data. An MCP server is a small program that exposes tools (like "read a file" or "create a GitHub issue") that any MCP-compatible agent can call.

MCP in one paragraph

Before MCP, every AI app had to write its own integration for every tool. MCP standardizes that connection: a tool author writes one MCP server, and it works in Claude Code, Claude Desktop, Cursor, Codex, VS Code, and every other MCP client. Anthropic introduced MCP in November 2024 as an open protocol, and it has since been adopted across the industry.

What an MCP server actually is

An MCP server is a regular program. The agent (the MCP client) starts it and talks to it with JSON-RPC 2.0 messages. A server can offer three kinds of things:

Tools: functions the agent can call, such as read_file, search_issues, or query_database.
Resources: data the agent can read, such as files, documents, or records.
Prompts: reusable prompt templates the user can pick from.

How the agent and the server talk

Local MCP servers usually run over stdio: the client launches the server as a subprocess and they exchange one JSON message per line on stdin and stdout. Remote servers use Streamable HTTP instead. A typical session looks like this: initialize, tools/list to discover the tools, then a series of tools/call requests while the agent works.

A tools/call request and its response

→ {"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"read_text_file","arguments":{"path":"notes/todo.txt"}}}
← {"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":"buy milk"}]}}

Seeing it for yourself

The fastest way to understand MCP is to watch real traffic. RelayOrb is a free, open-source recorder that sits between your agent and an MCP server and saves every message. Wrap any server with relayorb record, use your agent normally, then run relayorb show to see each call, its arguments, the answer, and how long it took.

Record an MCP server and inspect the session

relayorb record --name fs -- npx -y @modelcontextprotocol/server-filesystem ~/notes
relayorb show fs

Install RelayOrb

Free and open source (Apache-2.0). A single binary for macOS, Linux, and Windows. No account, no cloud, no telemetry.

macOS / Linux

curl -fsSL https://relayorb.com/install.sh | sh

Windows and other options: GitHub Releases.

FAQ

What does MCP stand for?: Model Context Protocol. It is an open protocol for connecting AI models and agents to external tools and data sources.
Is an MCP server a web server?: Not necessarily. Most local MCP servers are command-line programs that the agent starts and talks to over stdin/stdout (the stdio transport). Remote MCP servers use HTTP.
Which apps support MCP?: Claude Code, Claude Desktop, Cursor, OpenAI Codex, VS Code, Windsurf, and many other agents and IDEs support MCP servers.
How can I see what an MCP server is doing?: Put a recorder like RelayOrb in front of it. RelayOrb relays the traffic unchanged and saves every request and response so you can inspect, replay, or test it later.