How to add an MCP server to an AI assistant

A remote server: add its URL

Assistants that accept remote MCP servers usually do so from their settings, under a name such as connectors, integrations, apps or tools; on team and business plans an administrator may need to allow it first. You add a server by giving it a name and its endpoint URL — the Streamable HTTP or SSE address on the server's page.

If the server requires sign-in, the assistant opens the service's own sign-in page and keeps the access it grants. If the registry entry declares a header, such as Authorization with an API key, the client needs a field for custom headers. Once connected, the assistant sees the server's tools and can call them in a conversation, usually asking before an action that changes something.

A local server: add its command

Desktop apps, code editors and command-line agents that run local servers read a configuration listing each server's command, arguments and environment variables. A common form is a JSON file with an mcpServers object; command-line agents often have a command that writes the same entry for you. The client starts each program when it needs it and stops it when it quits.

Example (placeholder names):

{
  "mcpServers": {
    "example": {
      "command": "npx",
      "args": ["-y", "example-mcp-server@1.2.0"],
      "env": { "EXAMPLE_API_KEY": "<your key>" }
    }
  }
}

When it does not connect

Most clients show each server's status and its list of tools. If a local server fails to start, run its command in a terminal: a missing runtime (Node.js, uv, Docker) or a missing environment variable is the usual cause. If a remote server is refused, check that the client supports its transport and whether it needs a header or a sign-in.