What is an MCP server?

What MCP is

MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems. With MCP, apps can reach data sources (for example local files or databases), tools (for example search or calculators) and workflows (for example specialized prompts). Source

Official docs compare MCP to a USB-C port for AI applications: a shared way to plug outside systems into many hosts, instead of a custom integration for each one. Source

Host, client and server

MCP uses a client-server setup. Source

• MCP hostThe AI application that coordinates connections, such as Claude Desktop, Claude Code, Cursor or VS Code. Source
• MCP clientA component inside the host. The host creates one client per server, and that client keeps the connection to that server. Source
• MCP serverA program that provides context to clients. "Server" here means the program that serves context, whether it runs on your computer or on the network. Source

Tools, resources and prompts

Servers expose capabilities through three building blocks. Source

• ToolsFunctions the model can call. The model decides when to use them. Tools can write to databases, call APIs, change files or trigger other logic. Clients discover them with tools/list and run them with tools/call. Source
• ResourcesPassive, read-only data for context (file contents, schemas, docs). Each resource has a URI. The application chooses what to fetch and how to use it. Source
• PromptsPre-built instruction templates. The user picks them; they are not meant to fire on their own. Source

A server can offer any mix of these. Many useful servers expose only tools.

Local (stdio) vs remote (Streamable HTTP)

MCP messages use JSON-RPC. The protocol defines two standard transports. Source Source

• stdioThe client launches the server as a subprocess on the same machine and exchanges newline-delimited JSON-RPC messages on stdin and stdout. Official guidance says clients should support stdio whenever possible. Local MCP servers typically serve one client this way. Source Source
• Streamable HTTPThe server runs as its own process and exposes an HTTP endpoint (for example https://example.com/mcp). Clients send JSON-RPC over HTTP POST; the server may reply with JSON or with Server-Sent Events. Remote servers typically serve many clients this way. Source Source

Custom transports are allowed if they keep the JSON-RPC message format. Source

How apps add a server

Each host has its own settings UI or config file. In broad terms:

• Local serverYou give the app a command (and args/env) so it can start the program. Claude Desktop stores this in claude_desktop_config.json. Source Claude Code, Cursor and VS Code each have their own MCP config paths. Source Source Source
• Remote serverYou give the app a URL and complete any sign-in the server asks for. Claude Desktop's docs cover adding remote connectors in Settings. Source

Our MCP config builder writes the local config shape for Claude Desktop, Claude Code, Cursor and VS Code. Related guides: Claude Code memory, one memory across AI apps, and compare memory tools.

Security tips

An MCP server is ordinary software with the privileges you give it. Official local-server guidance treats a third-party local server as untrusted code with exactly those privileges. Source

• Only connect servers you trustKnow the publisher and where the package comes from before you install. Prefer pinned versions. Remove servers you no longer use. Source Source
• Review what tools can doRead each tool's name and description. Tool text is instruction the model sees; skip servers whose descriptions ask for unrelated actions. Source
• Limit accessGive each server only the credentials and folders it needs. Isolation (for example a container) shrinks what a bad server can reach. Source

Example: a memory server

A memory server is an MCP server whose tools save and look up notes outside any one chat. FlowState is one example: a free, open-source local MCP server that keeps project memory (to-dos, decisions, changes, problems and lessons) in a SQLite file on your computer. Connected apps call tools such as start_here, log_work and search. It is one kind of MCP server among many; file, database and API servers are other common kinds.

See How to give Claude Code memory between sessions and One memory for Claude, ChatGPT, Cursor and Grok, or compare FlowState with other memory tools.

Sources

1. What MCP is
2. Host, client and server; transports overview
3. Tools, resources and prompts
4. stdio and Streamable HTTP transports
5. Local server security tips
6. Connecting local servers (Claude Desktop example)
7. Connecting remote servers
8. Claude Code MCP docs
9. Cursor MCP docs
10. VS Code MCP configuration

Questions

• What is MCP?The Model Context Protocol is an open standard for connecting AI applications to external systems such as data sources, tools and workflows.
• What is an MCP server?A program that provides context to MCP clients. It can expose tools the model may call, resources the app can read, and prompts the user can invoke.
• How is a server different from a client or host?The host is the AI application (for example Claude Desktop or an IDE). For each server it connects to, the host creates one MCP client that keeps that connection. The server is the program that supplies the context.
• What are tools, resources and prompts?Tools are functions the model can call. Resources are read-only data the application can fetch for context. Prompts are reusable templates the user chooses to run.
• Local or remote?Local servers usually use the stdio transport: the client starts the server as a subprocess on the same computer. Remote servers use Streamable HTTP over a URL so the client and server can run on different machines.
• Is FlowState an MCP server?Yes. FlowState is a local MCP memory server. AI apps connect to it and use its tools to save and look up project notes such as decisions and to-dos.