Remote MCP Servers: How They Work and How to Connect One

The Model Context Protocol defines how an AI application (the client) discovers and calls tools on a server. The first MCP servers ran locally: your editor started a program on your machine and talked to it over standard input and output. A remote MCP server is the same idea delivered as a web service. It already runs somewhere on the internet, and the client connects to its URL.

That changes who does the work. With a remote server there is nothing to install, no runtime to keep up to date, and one deployment serves every user. In exchange the server has to handle what a local process never worried about: authenticating each user, keeping their data apart, and staying available.

MCP messages are JSON-RPC either way. The transport is what carries them, and it decides whether a server can be remote.

Streamable HTTP replaced the earlier HTTP+SSE transport. A server can still stream responses when a tool takes a while, but everything starts with a POST to one endpoint. Our short note on Streamable HTTP or stdio covers what that means for MCPBackend specifically.

Because anyone can reach a public URL, a remote server that touches real accounts has to know who is calling. The MCP authorization spec builds on OAuth 2.1, and a typical first connection looks like this:

1. 1

   The client calls the server and gets a 401

   The response points to the server’s OAuth metadata, which lists its authorization and token endpoints.

2. 2

   The client registers itself

   Many servers support dynamic client registration, so the client gets a client ID without you creating one by hand.

3. 3

   You approve access in the browser

   The client opens the server’s consent page. You sign in there; your password or code never passes through the AI tool.

4. 4

   The client exchanges the code for tokens

   PKCE protects the exchange. The client stores an access token and a refresh token and sends the access token on each request.

5. 5

   Tokens refresh quietly

   When the access token expires, the client uses the refresh token instead of asking you to sign in again.

MCPBackend follows this flow. If you are already signed in to the dashboard you approve straight away; otherwise it emails a four-digit code and returns you to the consent page.

Every client needs the same two things, a name and the server URL. These examples use MCPBackend; swap in any other remote server’s URL.

Claude Code:

claude mcp add --transport http mcpbackend https://mcp.mcpbackend.com/mcp

Cursor, in ~/.cursor/mcp.json or a project’s .cursor/mcp.json:

{
  "mcpServers": {
    "mcpbackend": { "url": "https://mcp.mcpbackend.com/mcp" }
  }
}

Claude and Claude Desktop: Settings → Connectors → Add custom connector, paste the URL and click Connect. Other clients, including ChatGPT’s connector settings, have their own forms and plan requirements, so check that the client supports remote servers with OAuth before you start. The MCP config generator produces the right snippet for several clients.

Once connected, ask the agent to list its tools. If they show up, the connection and sign-in worked.

If you are exposing your own product to AI clients, you will write and deploy a server. The common starting points are:

• TypeScript SDK. The official @modelcontextprotocol/sdk package includes a Streamable HTTP server transport you can mount in Node frameworks such as Express or Hono.
• Python and FastMCP. The official Python SDK’s FastMCP interface can serve over Streamable HTTP instead of stdio, which makes it straightforward to run behind any Python web host.
• Cloudflare Workers. The Cloudflare Agents SDK provides an MCP server class and an OAuth provider library, so a server can run on the edge without managing machines. MCPBackend’s own MCP server runs on Workers.

Whichever you pick, the hard parts are the same: OAuth and token storage, isolating each user’s data, sensible timeouts for slow tools, rate limits, and logging that never records secrets. Budget for those before the tool code itself.

You may not need a server of your own at all. If the goal is a backend for an app you are building with an AI editor, connecting a hosted MCP database gives the agent tables, auth and an API without anything to deploy. See what an MCP backend is for how that model works.

Most connection problems fall into a few patterns:

The step-by-step fixes are in the developer library: unauthorized after connecting, OAuth consent loop, tools not visible and tool timeouts. The full MCP and agent workflows topic has more.