mcpmcp serversclaudechatgptcursorconnectorssmall business

Remote MCP servers: what they are and how to add one

A hosted server takes an address and a sign-in, with nothing to install or update. The difference from local ones, and the checks worth doing before you allow one.

Updated 13 minutes
Two people collaborate at a desk with a laptop and a QR code card for remote MCP servers.
Two people collaborate at a desk with a laptop and a QR code card for remote MCP servers.

Remote MCP servers are the ones you add by pasting an HTTPS address and signing in, with no install on your computer and no config file to keep up to date.

That one difference decides almost everything else: who can set it up, whether it works on your phone, and what you are agreeing to when you press Allow. Local servers still exist and still have a job, but for most people the hosted kind is the only one worth the trouble.

This piece is about the hosted kind itself. For which servers are worth connecting in the first place, the 12 best MCP servers for a small team has the roster.

How do remote MCP servers differ from local ones?#

Two people sitting at a table discussing with a QR code card between them in an office setting.
Remote means an address and a sign-in page, nothing to install.

A remote server runs on the vendor's own machines and is reached by an HTTPS address. A local one runs on your computer, needs an install, and usually needs a key pasted into a config file by hand.

RemoteLocal
How you add itPaste an address, or press Connect in a directoryInstall a package, edit a config file
Sign-inThe vendor's own page, with a permission listOften an API key pasted by hand
Works on your phoneYes, for hosted assistantsNo
UpdatesThe vendor updates it and you notice nothingYou update it
Your filesNot reachableReachable, which is the point of it
Best forEveryoneDevelopers, and work with files on one machine

The row that surprises people is the last one. A hosted server cannot touch the files on your laptop, because it never runs there. That is a limit when you want an assistant to rename a folder of images, and a relief the rest of the time.

How can you tell a server is remote?#

Woman using a tablet in a restaurant with a QR code on a wooden stand nearby.
A printed code an assistant can re-point later is the one worth printing in quantity.

Look at what the vendor asks you for. An address that starts with https and a sign-in button means hosted. A command to run, a package to install, or a key to paste into a file means local.

Three signs you are looking at a hosted one:

  • It gives you a URL, not an install command. That address is the whole setup.
  • It appears in a connector directory inside Claude or ChatGPT, where a Connect button does the work.
  • It signs you in on its own website, on the vendor's real domain, and shows you a list of what the assistant will be allowed to do.

If a setup page asks you to paste an API key into the chat window, that is neither of the two done properly. Keys belong in a file on your own machine, or better, nowhere at all.

Where do you paste a server address?#

Man using a laptop at a desk with a QR code card, glasses, and a steaming mug nearby.
Own domain, listed permissions, a place to disconnect: the three checks.

Every assistant keeps this in a different place, but the shape is the same everywhere: find where connectors live, paste the address, then sign in on the vendor's page and press Allow.

AssistantWhereSign-in
ClaudeCustomize, then Connectors, then Add custom connector (or Connect in the directory)Vendor page, Allow
ChatGPTDeveloper mode on, then Settings, Plugins, the plus buttonVendor page, Allow
CursorCursor Settings, then MCP, one entry with URLBrowser, Allow
Claude Codeclaude mcp add --transport http name address, then /mcpBrowser, Allow
VS CodeMCP: Add Server, or .vscode/mcp.json with type httpBrowser, Allow
Windsurfmcp_config.json with serverUrlBrowser, Allow
Codex CLIcodex mcp add name --url address, then codex mcp loginBrowser, Allow
Gemini CLIgemini mcp add --transport http name address, then /mcp authBrowser, Allow

The assistants page has the same rows with pictures of each screen. Working in a terminal rather than a browser changes the job enough to be worth its own read: Claude Code MCP servers for marketers covers that side.

Three checks before a hosted server gets access#

Two people reviewing documents at a table with a QR code card displayed.
Verifying server access before connection.

Check the domain, the permission list and the way out, in that order. All three are visible on the sign-in page before you press anything.

  • Domain. A payments server lives on the payments company's own domain. An address on an unrelated host is a reason to stop.
  • Permissions. The list on the sign-in page is the contract. A job that only reads should not be asking to write.
  • Undo. The assistant's settings can remove any server, and most vendors also list connected apps on their own account page, so you can cut it from either end.

Test a new connection with a harmless read-only question first. "How many open invoices are there" tells you the connection works without changing a single record, and you only move on to create and update once the answers look right.

When is a local server still the right call?#

Two people sit at a table discussing with a QR code card and plants nearby.
Discussing the advantages of remote servers for small teams.

Local wins when the job is on your own machine. Files in a folder, a database running on your laptop, a tool that only ships as a package: none of those can be reached by a server running somewhere else.

For everything else, hosted is the easier answer for a small team. Nothing to install, nothing to keep updated, and the same connection follows your account to your phone. A manager checking sales from home does not want a config file in the way.

If a service offers both, take the hosted one unless you specifically need the machine access. You can always add the local version later for the one job that needs it.

A QR code MCP server is a good first hosted connection to try, because the result is something you can hold: ask for a code, print it, and change where it points a month later without reprinting. QRFLOW.codes does that from any assistant, and the address is the whole setup.

https://qrflow.codes/mcp

Questions people ask#

Are remote MCP servers free?#

Adding one costs nothing on the assistant's side. The vendor keeps its own pricing, and a server usually does what your existing plan with that vendor allows, nothing more.

Do I need Developer mode?#

In ChatGPT, yes, to add a server by address, though directory apps do not need it. Claude's free plan can add one custom connector and the paid plans add more. Editors and command-line tools have no plan gate for this at all.

Can one assistant use several at once?#

Yes. "Make a QR code for the new page and post it in the team channel" uses two servers in one request, as long as both are connected and both signed in.

What if a server wants an API key pasted into the chat?#

Do not paste it. A properly built hosted server signs you in on its own page and never needs a key in the conversation. A key in a chat log is a key you cannot take back.

Is a remote server the same as a connector or a plugin?#

Yes, those are three names for one thing. The vendor runs an MCP server, Claude shows it as a connector, and ChatGPT calls it a plugin or an app.

Start with the two tools you open every day, add them as hosted connections, and ask read-only questions for a week before you let anything write. If you only wanted the QR code, the free QR code generator makes one without any of this.

Your turn

Ready to make your own QR code?

Free QR codes rendered as you type, no signup to try. Premium makes them dynamic for $4 a month.