An MCP connector is a remote MCP server, reachable over HTTP, that an AI client can attach to by URL so the model can call the tools the server exposes. The word "connector" is what Anthropic's products and the Claude API call a hosted MCP server once it is wired in; the protocol itself only has servers and clients. In practice, every connector is an MCP server, but not every MCP server is offered as a connector.
Connector vs server vs client vs plug-in
| Term | What it is | Where it runs | Example | |---|---|---|---| | MCP server | A process that exposes tools, resources and prompts over the Model Context Protocol | Locally over stdio, or remotely over HTTP | Context7, GitHub MCP Server | | MCP connector | A remote MCP server presented to a product as something you connect by URL | On the server provider's infrastructure | The same GitHub server added to Claude.ai as a connector | | MCP client | The application that connects to servers and lets the model call their tools | Wherever the agent runs | Claude Code, Cursor, Claude Desktop, Claude.ai | | Plug-in | A product-specific extension format that is not MCP | Inside one product | ChatGPT plug-ins (retired), IDE extensions |
The MCP connector in the Claude API
Anthropic's Messages API accepts remote MCP servers in the request itself, documented as the MCP connector feature, so your application does not have to implement an MCP client. You pass the server URL in an mcp_servers array and the model calls its tools during the response:
{
"model": "claude-sonnet-5",
"max_tokens": 1024,
"messages": [{ "role": "user", "content": "What changed in the last release of vercel/next.js?" }],
"mcp_servers": [
{
"type": "url",
"url": "https://api.githubcopilot.com/mcp/",
"name": "github",
"authorization_token": "YOUR_TOKEN"
}
]
}
The server must speak the HTTP transport; a stdio-only server cannot be used this way. Check the current request format in Anthropic's API reference before shipping, since the beta header and field names have changed between releases.
MCP connectors in Claude.ai and Claude Desktop
In Claude.ai, connectors live under Settings, then Connectors. Add a custom connector by pasting the remote server's URL; the server handles sign-in with OAuth if it needs it. Claude Desktop accepts the same remote connectors and also runs local servers from claude_desktop_config.json. Each MCPVault listing shows which clients a server works with in the "Works with" panel, and the hosted endpoint where the project publishes one.
Frequently asked questions
What is an MCP connector?
A remote MCP server that a client attaches to by URL. It exposes tools the model can call, the same way a local MCP server does, without anything installed on your machine.
What does MCP stand for?
Model Context Protocol, the open standard introduced by Anthropic in November 2024 for connecting AI models to tools and data. The What is an MCP server guide covers the protocol itself.
What is an MCP plug-in?
MCP has servers and clients, no plug-ins. "Plug-in" usually refers to a product's own extension system, such as retired ChatGPT plug-ins or IDE extensions. Some of those now wrap an MCP server underneath.
How do you make an MCP connector?
Build an MCP server with the HTTP transport (the official SDKs for TypeScript and Python include it, and FastMCP does it in a few lines for Python), host it on a public URL, add OAuth if it touches user data, and publish the URL. Then list it on MCPVault so clients can find it.