What is MCP in simple terms?
The protocol docs compare MCP to a USB-C port. Before it, every AI app connected to every service in its own way. With MCP a service describes its capabilities once, and any MCP client can use them: Claude Code, Codex, Cursor, VS Code, ChatGPT.
An example from the same docs: Claude Code builds a web app from a Figma design because a Figma MCP server is connected. Without a server the agent sees only your code. With one, it can read a ticket from your tracker, query a database or open a page in a browser.
Who created MCP?
Anthropic built it; David Soria Parra and Justin Spahr-Summers are credited as its creators. Anthropic open-sourced it on 25 November 2024. The spec, SDKs and reference servers live in the open modelcontextprotocol organization on GitHub, and competitors support it too: OpenAI, GitHub, Google and Cursor.
How MCP works: host, client, server
- Host: the AI application, such as Claude Code or VS Code.
- Client: a component inside the host. The host creates one client per server, each with its own connection.
- Server: a program that provides context, running on your machine or somewhere on the network.
A server offers three kinds of capabilities:
PrimitiveWhat it isExampleToolsfunctions the model can callrun a SQL query, open an issueResourcesdata for contexta database schema, file contentsPromptsreusable templatesa ready-made debugging flowMessages are JSON-RPC 2.0. There are two transports. stdio: the host launches the server as a local process and talks over standard input and output, a “local” server. Streamable HTTP: the server runs remotely and is reached over HTTP with a token or OAuth, a “remote” server.
What is the difference between MCP and an API?
An API belongs to one service: GitHub has one, Postgres has another, and each needs its own integration code. MCP is a protocol on top of those APIs. A GitHub MCP server calls the GitHub API internally but exposes tools in one shape that any MCP client understands.
The second difference is self-description. The client asks the server for its tool list (tools/list) and gets a name, description and JSON schema for every tool. The model reads those descriptions and decides what to call. With a plain API that knowledge has to be hard-coded or stuffed into a prompt.
Bottom line: writing a program that talks to one service? Use the API. Want an AI agent to use the service on its own? You need an MCP server.
Which MCP servers exist and where to find them
There are many catalogs, and their numbers disagree because of duplicates. On 24 October 2026 Glama listed 109,544 servers, the Smithery registry API reported 20,773, and the mcp.so sitemap had 19,194 server pages. The most-starred GitHub list, punkpeye/awesome-mcp-servers, had 96,811 stars and 4,631 entries.
Common categories: databases, browser automation and scraping, GitHub and git, files, issue trackers, documentation. Servers we have checked for license and freshness are in the MCP section and in our picks for databases and browsers.
On safety: a server acts on your behalf. Claude Code asks for approval before using servers from a project .mcp.json, and Cursor asks before tool calls by default. Keep those checks on for any server that can write.
How to connect MCP to Claude Code, Cursor and ChatGPT
In Claude Code you add a server with claude mcp add, in Codex with codex mcp add, in Cursor through .cursor/mcp.json. Config formats differ between agents; the command table for eight clients is in how to add an MCP server.
For ChatGPT, the OpenAI developer docs say you add a remote server under Plugins: the plus button, then Add custom MCP server, then the server URL and auth details (OAuth recommended). That page does not say which ChatGPT plans include it. In the OpenAI API, a server is attached as a tool of type mcp with a server_url.
How do you build your own MCP server?
Start with the official docs: they include a Build servers guide and SDKs for several languages. A minimal server describes one tool (name, description, JSON input schema) plus the function that runs it. Test it with MCP Inspector, from the same GitHub organization, before you connect it to an agent.
The part that matters most is the tool descriptions. The model picks a tool by its description, just as it picks a skill. If a tool never gets called, rewrite the description before you touch the code. When to use MCP and when a skill is enough is covered in skills vs MCP.
