MCP clients: which AI apps connect to remote servers
An MCP client is the AI application that connects to an MCP server, discovers the tools it advertises, and calls them during a conversation or task. Claude and ChatGPT are the best-known clients; coding tools, automation platforms and desktop assistants have added support as well.
The distinction that matters when choosing one is local versus remote. A client that only supports local servers runs a process on your machine. A client that supports remote servers connects to a URL over HTTPS, which is what makes a hosted, multi-user server usable at all.
What a client needs to support to connect to a remote server
Three things, and a client missing any of them cannot connect no matter how the server is configured.
- Remote server configuration — the ability to add a server by URL rather than by local command
- An authentication path — either an OAuth flow, or the ability to send a fixed credential as a request header
- HTTP transport — servers that answer over HTTPS rather than standard input and output
Zylx requires exactly this and nothing exotic: a single HTTPS endpoint, authenticated by OAuth or by an Authorization header. Any client meeting the three points above can connect using the same server URL as every other client.
Verified against Zylx
Two clients have been confirmed end to end against the Zylx MCP server. These are stated as verified because the connection, the authorization flow and the tool calls were exercised, not because a specification implies they should work.
| Client | How it connects | Authentication | Setup guide |
|---|---|---|---|
| Claude | Custom connector, by server URL | OAuth, or a token in a request header | /mcp/claude |
| ChatGPT | Custom connector behind developer mode | OAuth | /mcp/chatgpt |
The wider client landscape
Beyond those two, a growing set of applications act as MCP clients — coding assistants such as Cursor, VS Code with Copilot, and Claude Code; desktop and CLI assistants; and automation platforms including n8n and Zapier that expose MCP as a node or step. Gemini has added MCP support through its CLI and enterprise tooling.
Two caveats, stated plainly. First, support in this space changes month to month, so treat any list — including this one — as a pointer to each vendor's current documentation rather than a specification. Second, these clients have not been tested against Zylx. They implement the same protocol and Zylx exposes a standard endpoint, so a connection should work; that is an expectation, not a verification, and we would rather say so.
It is also common for vendors to gate MCP behind paid plans, so check plan availability before assuming a client can add custom servers at all.
What differs between clients
The protocol is the same everywhere. What varies is the product wrapped around it, and those differences decide most of the friction you will actually hit.
| Dimension | Why it matters |
|---|---|
| Local vs remote servers | Determines whether a hosted server is usable at all |
| OAuth vs header credentials | Decides whether actions tie to a person or to a shared secret |
| Plan gating | Custom servers are frequently a paid-plan feature |
| Per-conversation control | Whether you can enable a server for one chat rather than all |
| Tool-level permissions | Whether individual tools can be blocked or require approval |
| Where the request originates | Some clients connect from vendor infrastructure, not your machine — which changes what the server must allow |
That last row catches people out. When a client connects from the vendor's cloud rather than your laptop, a server behind a VPN or a local tunnel is unreachable even though it works in your browser.
Frequently asked questions
What is an MCP client?
An MCP client is the AI application that connects to an MCP server, discovers the tools it advertises, and calls them while working. The server holds the data or capability; the client decides which tools to call and writes the answer.
What is the difference between an MCP client and an MCP server?
The server exposes named tools and enforces who may call them. The client is the AI app that connects to it. One client can connect to many servers, and one server can serve many clients.
Which AI apps support MCP?
Claude and ChatGPT both support remote MCP servers as custom connectors, and a growing set of coding assistants, desktop tools and automation platforms have added support. Availability and plan requirements change frequently, so check the current documentation for the client you intend to use.
Can one MCP server serve several different AI apps?
Yes, and that is much of the point of a standard. A server built once can be authorized by any compatible client, which is why context maintained behind an MCP server does not have to be rebuilt per assistant.
Do I need a paid plan to use MCP?
Often, on the client side — vendors commonly gate custom connectors behind paid tiers. The server may have its own requirements too; Zylx includes MCP access from the Pro plan and above.
Has Zylx been tested with clients other than Claude and ChatGPT?
No. Those two are verified end to end. Zylx exposes a standard remote MCP endpoint authenticated by OAuth or a bearer header, so other compatible clients should connect using the same URL and credential, but we have not tested them and do not claim them as verified.
One server, every client you authorize
Build the context once. Connect whichever assistant your team actually uses.
Connect ClaudeConnect ChatGPTMCP securityWhat the clients are reading
Sources
- Model Context Protocol specification and documentation
- Anthropic: third-party connectors with remote MCP
- OpenAI: building MCP servers
Zylx product details verified against the live implementation on 2026-08-09.