A remote MCP server is a Model Context Protocol server that runs on someone else's machine and answers at a web address, so your AI app connects to it over HTTPS instead of launching it on your computer. You paste a URL into Claude, ChatGPT, Claude Code or Cursor, sign in or add a key, and the tools appear. Nothing gets installed locally.

What is a remote MCP server, in plain terms?

MCP is the standard way an AI app talks to outside tools and data. The AI app is the client. The program that exposes tools such as "run a query" or "list my apps" is the server. Where that server runs is the only thing "remote" changes.

A local server starts on your own computer when the client launches it. A remote server is already running somewhere else. It listens at an endpoint such as https://example.com/mcp, and any client that knows the address and has permission can use it. People also call this a hosted MCP server, a cloud MCP server or an MCP server over HTTP. Those names point at the same idea.

The practical effect is large for people who are not developers. Local servers usually need Node.js, Python or Docker, a config file, and your laptop awake. A remote server needs a link. It also works from places a local process cannot reach, such as the Claude.ai website or the ChatGPT web app.

Remote vs local MCP server: what actually differs?

The two types differ in where the code runs, how the client reaches it, and how it proves who you are. This table sums it up.

Local server (stdio)Remote server (HTTP)

Where it runs

On your computer, started by the client

On a server on the internet

How the client connects

Standard input and output of a child process

HTTPS requests to one URL

Setup

Install a runtime or package, edit a config file

Paste a URL, then sign in or add a key

Works in web and phone apps

No

Yes, where the client supports remote servers

Authentication

Usually environment variables on your machine

OAuth or API keys sent with each request

Who updates it

You

The operator of the server

Sharing with a team

Each person repeats the setup

Everyone uses the same URL

Main risk

Full access to your machine and files

Exposed endpoint and the data you send to it

Neither type is better in every case. A local server suits a folder of files you never want to leave your machine. A remote one suits anything you want to reach from several apps or share with other people.

How does a remote MCP server connect: Streamable HTTP and legacy SSE?

MCP defines how messages travel, and the current standard transports are stdio and Streamable HTTP. The spec's transports page describes Streamable HTTP as one endpoint that accepts POST requests from the client and can reply with a single JSON response or open a stream of server-sent events when it has more to say. Our entry on Streamable HTTP covers the details.

Older servers use the legacy HTTP+SSE transport, which needed two endpoints: one to listen, one to send. The spec replaced it with Streamable HTTP in the 2025-03-26 revision and keeps a backwards-compatibility section for clients that still meet old servers. If a vendor tells you to use an "SSE URL", you are looking at the older style. Prefer the Streamable HTTP URL when both are offered.

The spec also tells servers to check the Origin header on incoming connections to prevent DNS rebinding attacks. Server operators handle this; you don't configure it.

How do remote MCP servers handle authentication?

Anyone who can reach a URL can try to call it, so every serious remote server asks for proof of identity. Two methods are common. Our MCP authentication entry goes deeper.

  • OAuth. The client sends you to a sign-in page, you approve access, and the client receives a token. The MCP authorization spec is built on OAuth 2.1. Pick this when the server acts on your account at a service such as a code host or an issue tracker, because you can revoke it there later.
  • API keys or bearer tokens. You copy a secret and the client sends it in an Authorization header. This is simpler, but the key works for whoever holds it, so treat it like a password and rotate it if it leaks.

Some web connectors cannot send custom headers. In that case a few services accept the key inside the URL path. It works, but a URL ends up in logs and screenshots more easily than a header does, so use the header whenever the client allows it.

Which AI clients support remote MCP servers?

Support has spread quickly, and the details change, so check each vendor's page before you rely on it. As of October 2026, these clients accept a remote URL:

Our list of MCP clients tracks more apps. A client that only supports stdio can still reach a remote server through a small bridge program, but that brings back the local install you were trying to avoid.

How do you get a remote MCP server?

There are three routes. They trade effort against control.

1. Use a vendor's official remote server

Many products now publish their own. Linear documents https://mcp.linear.app/mcp and a read-only variant in its MCP docs. Sentry documents https://mcp.sentry.dev/mcp in its MCP docs. You add the URL, sign in with OAuth, and you are done. The limit is that each vendor covers only its own product, so five tools mean five connections, each set up separately in every client.

2. Host your own

If you build servers, you can run one yourself. Cloudflare's guide to building a remote MCP server on Workers is one well-documented path, and any host that serves HTTPS can work. Expect to write session handling, authentication, rate limits and logging, then keep patching it. Our post on build vs buy for MCP hosting puts rough numbers on the work.

3. Use a hosted service such as MCPifex

MCPifex runs the real MCP server packages for you. You choose a server in the marketplace, save your own credentials, pick which tools are allowed and get an API key. Then you paste one URL, https://mcp.mcpifex.com/mcp, into your client. The five live servers are PostgreSQL, Google Trends, Google Search Console, App Store Connect and WhatsApp. Write tools start off, and you switch on only what you want. The Free plan includes 3 instances and needs no card. Start with the quickstart.

For a side-by-side view of the second and third routes, read hosted vs self-hosted MCP.

What are the risks of a remote MCP server?

A remote server moves your data and your credentials across the internet, so a few habits matter.

  • Trust the operator. Your prompts, tool arguments and results pass through their system. Use official servers or operators you can read about.
  • Limit the tools. A server that can read a database may also expose a tool that writes to it. Turn off anything you do not need. See how to control what an AI agent can reach.
  • Guard secrets. Prefer OAuth where offered, keep keys out of shared documents, and rotate any key that appears in a screenshot.
  • Expect prompt injection. Text returned by a tool can carry instructions meant to steer the model. Read prompt injection through MCP tools and keep write access narrow.
  • Check what you connect. Unknown public servers can change behavior after you add them. Our MCP security checklist lists what to review.

A gateway is a remote MCP server that sits in front of several others and gives your client one endpoint. A registry is a directory that lists servers but does not always run them. The Model Context Protocol is the standard behind all of it.

Quick answers

  • A remote MCP server is a Model Context Protocol server you reach at a URL over HTTPS instead of running it on your own computer. You add the address to an AI client such as Claude or ChatGPT, authenticate, and use its tools.

Sources

  1. Transports, MCP specification (2025-06-18)
  2. Authorization, MCP specification (2025-06-18)
  3. Get started with custom connectors using remote MCP, Claude Help Center
  4. Build a remote MCP server, Cloudflare Agents docs

Ready to try it?

Run hosted MCP servers behind one endpoint and choose exactly which tools your AI can use.

Start freeBrowse the marketplace

← Back to glossary