What is an MCP server?
An MCP server is a program that exposes tools an AI assistant can call. The Model Context Protocol is an open standard for that conversation: the server advertises what it can do, and the assistant decides when to use it. The server does the work; the model only chooses.
How does an MCP server actually work?
A client — Claude Code, Cursor, VS Code and others — starts or connects to the server and asks it to list its capabilities. The server answers with a set of tools, each with a name, a description and a JSON schema for its arguments.
From then on the model can call those tools. When it does, the client executes the call against the server and feeds the result back into the conversation. The model never runs code itself; it only picks which tool to invoke and with what arguments.
- Tools — functions the model can call, like `create_issue` or `run_query`.
- Resources — data the client can attach as context, like a file or a record.
- Prompts — reusable prompt templates the user can invoke directly.
Local or remote?
A local server runs on your machine as a subprocess and talks over stdio. It can reach your filesystem and your local network, and nothing leaves your machine unless the server sends it.
A remote server runs on the provider's infrastructure and you connect to it over HTTP. There is nothing to install, but your requests and any credentials you supply pass through a third party.
| Local | Remote | |
|---|---|---|
| Install | A runtime plus the package | Nothing — just a URL |
| Transport | stdio | streamable-http or SSE |
| Reaches your files | Yes | No |
| Credentials | Stay on your machine | Sent to the provider |
| Updates | You update it | Updated for you |
What an MCP server is not
It is not a plugin that changes how the model thinks, and it is not a way to fine-tune anything. It adds capability, not knowledge.
It is also not automatically safe. A server runs with whatever access you give it, and the model decides when to call it. Treat installing one like adding a dependency: read what it does, check who publishes it, and prefer servers whose scope is narrow.
How do I add one?
Every entry in this directory carries a ready-made command or config block for each client we can generate a verified one for. Pick your client, paste, restart it.
- Find the server you want and open its entry.
- Open the Install tab and choose your client.
- Copy the command or the JSON block into the file it names.
- Restart the client and confirm the tools appear.
常见问题
- Do I need an API key for every MCP server?
- No. Servers that only touch local resources usually need nothing. Servers that talk to a third-party service need whatever that service requires — normally an API key set as an environment variable. Each entry here states which of the two it is.
- Can I use MCP servers with any AI assistant?
- Only with clients that implement the protocol. Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Cline, Codex CLI, Zed, Goose and others do. A client that does not speak MCP cannot use these servers at all.
- How many MCP servers should I install?
- Fewer than you would expect. Every connected server's tool definitions are loaded into the model's context, so a dozen servers can consume a large slice of the context window before you have asked anything. Install what a given project actually needs.
继续阅读
- Local vs remote MCP serversLocal servers run on your machine and can reach your files. Remote servers run on someone else's and cannot. Which to choose, and what each one costs you.
- MCP server security: what to check before installingAn MCP server runs with the access you give it and the model decides when to call it. The practical checks that matter, in the order they matter.