Remote MCP Servers: What They Are and How to Connect One

Short answer: Remote MCP servers are MCP servers that run on someone else’s machine and connect to your AI app through a web URL, instead of a program on your own computer. Most are hosted by the vendor and you sign in with OAuth once. In Claude Code one command connects it: claude mcp add --transport http <name> <url>.

Remote servers are the easiest kind to start with, because there is nothing to install. Browse the Klyvoo MCP servers directory to see which servers offer a remote option, or read what MCP servers are first if the term is new.

Remote MCP servers at a glance

What is a remote MCP server?

A remote MCP server is an MCP server that your AI app reaches over the network at a URL, such as https://mcp.linear.app/mcp, rather than by launching it as a program on your own machine. The server does the same job as any other MCP server: it exposes tools, data and prompts to the AI app. Only the way you reach it changes.

The MCP specification defines two standard transports. With stdio, “the client launches the MCP server as a subprocess” and talks to it through standard input and output. With Streamable HTTP, the server “operates as an independent process that can handle multiple client connections” and uses HTTP POST and GET requests. A remote server is a server you reach over Streamable HTTP.

The spec says the server must provide “a single HTTP endpoint path” that handles both POST and GET, for example https://example.com/mcp. That one address is what you paste into your AI app.

How does a remote server work?

Your AI app sends each message to the server’s endpoint as an HTTP POST. The server can answer with one JSON reply or open a stream of events (Server-Sent Events, or SSE) when it has more to send, such as progress updates. The app can also open a GET connection so the server can send messages without being asked first.

Some servers hand out a session ID in an Mcp-Session-Id header at the start. If one is issued, the app sends it back on every later request. You never see any of this. It matters only when a connection drops and you wonder why.

What happened to SSE servers?

Before Streamable HTTP, MCP used a different transport called HTTP+SSE. The spec says Streamable HTTP replaces it, and calls the old one deprecated. The Claude Code docs say the same: the SSE transport “is deprecated. Use HTTP servers instead, where available.” You will still see older URLs that end in /sse, for example the deprecated Linear endpoint https://mcp.linear.app/sse. For anything new, use the /mcp address.

Newer versions of Claude Code try HTTP first and switch to SSE when a server does not accept it. Per the docs, that applies from version 2.1.265, so on older versions you may need to name the transport yourself.

How does sign-in work for a remote server?

Remote servers can hold your data, so most need you to prove who you are. The MCP spec makes authorization optional, and when a server uses it, the flow is based on OAuth 2.1. In plain terms:

1. Your app contacts the server without a token and the server answers “401 Unauthorized”.
2. The reply tells the app where to find the server’s sign-in service.
3. Your browser opens, you log in to the vendor and approve access.
4. The app receives an access token and sends it with each request in an Authorization: Bearer header.

The spec requires apps to use PKCE, a check that stops a stolen sign-in code from being reused. It also says tokens must not go in the URL query string and that servers must only accept tokens issued for them. Apps and servers should support dynamic client registration, which lets your app sign up with a new server automatically, so you do not have to create an OAuth app by hand.

Local stdio servers are different. The spec says they should not follow this flow and should “retrieve credentials from the environment” instead. That is why a local server usually asks for an API key in a config file, while a remote one opens a browser window.

Some vendors also offer a header route. Linear lets you pass an OAuth token or API key as Authorization: Bearer <yourtoken> and skip the interactive step. Notion offers tokens for headless use on its Business and Enterprise plans. Check the vendor’s page for which of these your plan allows.

How do you connect a remote MCP server?

The steps depend on your AI app. These are the general ones, with Claude Code as the worked example.

1. Find the endpoint. Open the server’s page on Klyvoo or the vendor’s docs and copy the URL. Look for the address ending in /mcp.
2. Add it. In Claude Code, the pattern from the docs is:

bash claude mcp add --transport http <name> <url>

For Notion’s hosted server, the docs give this exact command:

bash claude mcp add --transport http notion https://mcp.notion.com/mcp

1. Sign in. Start Claude Code and run /mcp, then follow the browser steps. Claude Code stores and refreshes the tokens for you. If the browser does not open, copy the URL from the prompt and open it by hand.
2. Use a header if the server wants a token. Add --header "Authorization: Bearer your-token" to the command. You can repeat the flag for more headers.
3. Check it works. Run /mcp again and look for the server in the list, then ask Claude to do something small with it, like listing your projects.

To remove access later, open the /mcp menu and choose “Clear authentication.”

For a fuller walkthrough with scopes and more examples, see our guide to adding MCP servers to Claude Code. The other clients have their own guides: Claude on desktop and web, ChatGPT and OpenAI Codex. Claude Code itself is listed at /agents/claude-code/.

Which servers offer a remote option?

On the Klyvoo directory, 70 of the 100 servers list a remote option and 30 are local only. Three that show the range:

• GitHub MCP Server: both remote and local. GitHub’s hosted endpoint is https://api.githubcopilot.com/mcp/, with “no local setup or runtime required.” Adding /readonly to a toolset URL limits it to read access.
• Notion MCP: both. The remote server is hosted by Notion at https://mcp.notion.com/mcp and uses OAuth.
• Linear MCP: remote. The endpoint is https://mcp.linear.app/mcp, and a separate read-only address, https://mcp.linear.app/mcp/readonly, “only ever exposes read tools.”

Servers that touch your own machine tend to be local only, such as Playwright MCP for driving a browser and Filesystem MCP for reading your files. A hosted server cannot reach into your laptop, so those stay local. You can browse by job in the developer tools and docs, notes and projects categories.

When should you pick remote over local?

Pick remote when:

• The tool is a web service you already use, such as a project tracker or a code host.
• You do not want to install or update anything.
• You want sign-in handled by the vendor, with access you can revoke from the /mcp menu.

Pick local when:

• The server needs your files, your browser or something running on your machine.
• You want to see and control the code that runs.
• The data should not leave your computer.

If a client only supports local servers, a bridge can help. Notion and Linear both document npx -y mcp-remote <url> as a way to reach a remote server from a client that has no remote support. It runs a small local program that forwards to the URL, so you are still sending data to the remote server.

What can’t remote servers do, and what should you watch?

• They cannot see your machine. No local files, no local browser, no apps running on your computer.
• They depend on the vendor. If the vendor changes the URL, retires an endpoint or has an outage, your connection breaks.
• Your data travels. Whatever the model asks the server to read or write goes to the vendor’s systems. Read what permissions you approve at the sign-in screen.
• A token is a key. Anyone who has your bearer token has your access. Keep it out of shared config files and screenshots.
• Not every server is trustworthy. A remote server you did not choose can return anything it likes. Connect only servers from vendors you trust, and prefer a read-only endpoint when it exists.

The spec adds a warning for people who run servers: on the web, servers must check the Origin header to guard against DNS rebinding attacks, and local ones should bind to 127.0.0.1 rather than every network interface. If you build servers yourself, our guide to building your first MCP server is a good place to start, and MCP vs API explains why a standard plug beats a custom integration.

FAQ

What is a remote MCP server?

A remote MCP server runs somewhere else, usually on the vendor’s side. Your AI app connects to it by URL over Streamable HTTP. A local server runs on your own computer and talks to the app over stdio. Both give the AI the same kinds of tools and data.

How do I connect to a remote MCP server?

Copy the server’s endpoint URL and add it in your AI app. In Claude Code, run claude mcp add --transport http <name> <url>, then run /mcp to sign in. Other apps have a similar add-a-connector screen that asks for the same URL.

Do remote MCP servers need an API key?

Often not. Many use OAuth, which opens a browser so you can log in and approve access. Some also accept a token or API key in an Authorization: Bearer header, which is useful when there is no browser, such as on a server.

Is SSE still used for remote MCP servers?

Only for older servers. The spec replaced the HTTP+SSE transport with Streamable HTTP, and Claude Code’s docs call SSE deprecated. Streamable HTTP can still use SSE inside it to stream messages, which is a different thing.

Are remote MCP servers safe?

They are as safe as the vendor behind them and the permissions you approve. Use OAuth where offered, choose read-only endpoints when you only need to read, and revoke access in the /mcp menu when you stop using a server.

Are remote MCP servers free?

The server itself is often free but you need an account with the vendor. Klyvoo lists Linear MCP and Notion MCP as free with an account, and GitHub’s server as free with a GitHub account. Plans and limits belong to the vendor, so check their pricing page.

Ready to find one to connect? Open the Klyvoo MCP servers directory and pick a server for the job you need done.