What is a remote MCP server, and how agents use one to publish

A remote MCP server is a service on the internet that an AI agent can connect to and use through the Model Context Protocol. Instead of installing anything, you give your agent (Claude, ChatGPT, Cursor, and others) a URL. You sign in once in your browser, and from then on the agent can call the server's tools, such as "publish this document" or "read the comments on that document", as part of a normal conversation.

This page explains what that means, how it differs from a local MCP server, and what happens when an agent uses one, using Offprint's server as the running example.

MCP in one paragraph

The Model Context Protocol (MCP) is an open standard for connecting AI applications to outside tools and data. An MCP client is the app you talk to: Claude Desktop, ChatGPT, Claude Code, Cursor, Zed, and so on. An MCP server offers capabilities to that client, mainly tools (actions the agent can take, each with a name, a description, and typed inputs). The client tells the model which tools exist; when the model decides one would help, the client calls it and hands the result back to the model.

Because it's a shared standard, a server written once works in every client that supports MCP. That's why one Offprint address works in Claude, ChatGPT, Cursor, and the rest.

Local vs. remote servers

	Local server	Remote server
Where it runs	On your computer, started by the client	On the internet, run by the provider
How you add it	A command to run, often with a package install	A URL
Transport	Standard input/output with a local process	HTTP (Streamable HTTP)
Sign-in	Usually an API key in a config file	Usually OAuth in the browser
Works on	The one machine where it's installed	Any client and device where you add the URL
Good for	Files and tools on your machine	Shared services and data that live online

Local servers are great for things only your computer can reach, like your file system. A remote server is the right shape for a service: the data lives with the provider, every device sees the same thing, and there's nothing to install or update.

How sign-in works

Remote MCP servers typically use OAuth, the same standard behind "Sign in with Google". With Offprint, the flow looks like this:

1. You add https://offprint.io/mcp to your client.
2. The client asks the server what it needs. The server points to its authorization server and describes the access it can grant.
3. The client registers itself with Offprint automatically (OAuth dynamic client registration), so you don't create an app or copy a client ID.
4. Your browser opens Offprint's consent screen. You sign in, see which client is asking and what it wants (reading and writing your documents), and approve.
5. The client receives a token and uses it for every tool call. You can revoke it at any time in Offprint's settings, and the agent loses access immediately.

The important property: your password never goes to the agent, and you can see and revoke every agent you've connected.

What a tool call looks like

Once connected, the agent sees Offprint's tools, each described in plain language. When you say "publish this summary to Offprint", the model chooses the publish_document tool and fills in its inputs:

{
  "title": "Weekly status, week of Oct 5",
  "markdown": "## Summary\n\nCheckout v2 reached 50% of traffic…",
  "visibility": "public"
}

The server checks the token, stores the document as revision 1, and returns its address, which the agent shows you. Other tools work the same way: update_document writes a new revision, list_comments reads reviewers' feedback, and resolve_comment closes a thread. The Connect page lists all of them.

Why publishing fits MCP well

Agents produce a lot of writing that people need to read: reports, plans, summaries. Without a tool, that writing stays in the chat. A remote MCP server gives the agent a place to put it that:

• exists outside the conversation, at a link anyone can open;
• records who wrote what, because every call arrives with a known client and account, so each revision can be stamped with its author (more on provenance);
• lets people answer back, since the same server lets the agent read comments and revise.

Adding one to your client

Every client has a slightly different place to paste the URL. Step-by-step instructions for Claude Code, Claude Desktop, ChatGPT, Cursor, Windsurf, Zed, Cline, and Continue are on the Connect page, and the guides walk through each one in full.

Common questions

Is MCP specific to one AI company? No. It's an open protocol supported by many AI applications and developer tools.

Can a remote server see my whole conversation? No. It only receives what the agent sends in a tool call, such as the document you asked it to publish.

What if I stop trusting an agent? Revoke it in the server's settings. With Offprint, that's Settings → Connected agents.