MCP Supported Clients: Claude, ChatGPT, IDEs and More

MCP supported clients include Claude Code, Claude Desktop, claude.ai, ChatGPT, Codex, Cursor, VS Code with Copilot, Gemini CLI and n8n. Almost every major AI app can now connect to an MCP server. What differs is how: local or remote, which sign-in method, and which account you need. To use MCPifex, you want Streamable HTTP and a way to send your API key.

A quick example

Say you use Claude. That could mean three different things. In Claude Code, you'd run one terminal command to add a server. In Claude Desktop or claude.ai, you'd add a custom connector through the app's settings. Each has its own setup, and one working doesn't prove the others do.

So "does Claude support MCP?" is the wrong question. The right one is "does this exact app, on this account, connect to this server?" The table below answers it for the main clients.

MCP client support at a glance

This is a connection guide based on each vendor's documentation, checked in September 2026. "Local" means the app starts a server on your machine over stdio. "Remote" means it connects to a URL. It doesn't promise that every optional MCP feature works in every app.

New to the terms? The MCP client definition explains hosts, clients and servers.

Check the app, transport, sign-in and feature

1. Which app are you really using?

The same model shows up in many apps with different tools. Write down the actual app before you pick instructions. Browser products may also let you use a published integration but not add your own server URL. And a cloud coding agent doesn't share your laptop's network or config.

2. Does the transport match?

A stdio setup names a command to launch. A remote setup names a URL. Pasting a URL into a field that expects a command won't work. "HTTP" and "SSE" also aren't the same thing, so follow the server's documented transport and pick the matching client option.

MCPifex uses Streamable HTTP with the handshake-based MCP transport. Clients on the newest spec revision interoperate through its backward-compatibility rules, which we cover in what changed in the July 2026 spec.

3. Which sign-in can the client use?

Some clients send an Authorization header. Some run an OAuth flow. Some only take a URL. An OAuth-only client can't use an API key as if it were an OAuth secret.

MCPifex accepts the same key as a bearer header or as the first path segment of the URL, which helps clients that can't send headers. A URL with your key in it is a password, so keep it out of screenshots and shared notes. The API key docs explain both forms.

4. Does it support what you need?

Tool support doesn't automatically mean resources, prompts or interactive input work too. Test the feature you actually plan to use. For a reporting task, that's discovering one tool and running a small read with known input.

Picking a Claude connection

For Claude Code, follow the Claude Code guide. Check its config scope, since a server added to one project isn't available everywhere. For Claude Desktop, use the Desktop guide. For the browser, use the claude.ai connector guide.

Remote connectors send requests from Anthropic's servers, even in the Desktop app. A private address that works on your laptop can be unreachable to them. Anthropic lists custom connectors on Free, Pro, Max, Team and Enterprise, with limits on Free.

ChatGPT web versus Codex

ChatGPT web needs developer mode, which OpenAI lists for Pro, Plus, Business, Enterprise and Education accounts. Workspace policy can also affect it. Our ChatGPT guide walks through it.

Local MCP in the ChatGPT desktop app and Codex is a separate route, and ChatGPT web doesn't read your local Codex config. When you report a problem, name the surface. A working terminal setup says nothing about the browser one.

Editors and workflow tools

The Cursor guide and VS Code guide use different config formats. Copying a whole config between editors often fails even if both support the same transport. Carry over the URL and the sign-in method, then use the new app's own format.

For n8n, the MCP Client guide covers consuming tools. Exposing an n8n workflow as an MCP server is the opposite direction.

A five-step test for any client

1. Note the app version, account type, transport and sign-in method, without writing down secrets.
2. Confirm the server itself works first. In the MCPifex portal, Test connection runs a real health-check call.
3. Connect the client and compare its tool list with the tools you enabled.
4. Run one harmless, narrow read, and look at the raw result before trusting the model's summary.
5. Restart the client and run it again, to catch setups that only worked once.

Where MCPifex fits

We built MCPifex so one API key works across the clients above. Pick a server, choose its tools and connect. The quickstart shows the whole flow, and pricing covers instance limits. Plans don't change what a third-party client supports.