How to Connect a Remote MCP Server to Claude and ChatGPT
A practical guide to connecting a remote MCP server to Claude custom connectors and ChatGPT custom MCP servers, covering reachability, Streamable HTTP, authentication options, read and write tools, and troubleshooting.

Kelis Shekhaliya

TL;DR
To connect a remote MCP server to Claude or ChatGPT, make it publicly reachable, serve it over Streamable HTTP, and pick an authentication method each client supports. Claude adds it as a custom connector and supports OAuth, no sign-in, or fixed credentials in request headers. ChatGPT adds it as a custom MCP server plugin on the web and supports OAuth, no authentication, or a mix. OAuth is the most portable choice when you want both clients and per-user permissions.
To connect a remote MCP server to Claude or ChatGPT, you need three things in place: a server that is reachable over the public internet, a transport the client supports (Streamable HTTP is the current standard for remote servers), and an authentication method the client can actually use. Once those are ready, adding the server is mostly a matter of pasting its URL into the client's custom connector or custom MCP server screen and completing sign-in.
This guide is for SaaS founders, CTOs, and API teams who have an MCP server, or are about to have one, and want their customers or internal teams to use it from Claude and ChatGPT. It covers what each client expects, the setup steps, authentication choices, and the problems that most often stop a connection from working.
What connecting a remote MCP server means
A remote MCP server is an MCP endpoint hosted on infrastructure you run or a provider runs for you. An AI application acts as the MCP client: it connects to the endpoint, discovers the tools, resources, and prompts the server publishes, and calls those tools during a conversation.
Remote servers differ from local servers in one important way. A local server runs on the user's computer, usually as a command the client starts. A remote server is a URL. That makes it the practical choice for a SaaS product, because every customer connects to the same managed endpoint instead of installing and running code on their own machine.
Both Claude and ChatGPT add remote MCP servers by URL:
Claude calls them custom connectors. According to Anthropic's help center, custom connectors are available on Claude, Claude Desktop, and Cowork across Free, Pro, Max, Ultra, Team, and Enterprise plans, with Free users limited to one custom connector.
ChatGPT lets you add a custom MCP server on the web, which becomes a plugin you install and then use in conversations.
Before you connect: a readiness checklist
Most failed connections trace back to something that was not ready on the server side. Check these first.
1. The server is publicly reachable
Claude connects to custom connectors from Anthropic's cloud infrastructure, not from the user's device. Anthropic notes this is true even in Claude Desktop and Cowork. A server on a private corporate network, behind a VPN, or blocked by a firewall will not connect, even if you can open it from your own laptop. If the server must stay on a private network, Anthropic's guidance is to allowlist its published IP ranges.
ChatGPT also expects a reachable server URL, and its setup screen offers a Secure MCP Tunnel option as an alternative to a public URL.
2. The transport is Streamable HTTP
Streamable HTTP is the transport the MCP specification defines for remote servers. Anthropic's connector documentation recommends it and describes the older HTTP+SSE transport as legacy and being deprecated. OpenAI's documentation lists both SSE and streaming HTTP as supported for ChatGPT. If you are building or choosing a server today, Streamable HTTP is the safe default for both clients.
3. Authentication matches what the client supports
This is where Claude and ChatGPT differ the most, so decide it before you hand anyone a URL. The next section covers it in detail.
4. The tool surface is deliberate
Every tool you publish becomes something the model can choose and call. Expose a focused set of operations with clear names, descriptions, and input schemas, and keep internal or risky endpoints out. Our guide on choosing which API endpoints to expose as MCP tools walks through that review.
5. The server has been tested
Test discovery and tool calls with a tool such as MCP Inspector before you test inside Claude or ChatGPT. It is much easier to separate a server bug from a client configuration problem that way. See how to test an MCP server with MCP Inspector.
Authentication options in Claude and ChatGPT
Question | Claude custom connector | ChatGPT custom MCP server |
|---|---|---|
OAuth | Supported. Claude can use its published client identity, register automatically, or use your own OAuth client | Supported. ChatGPT can use static client credentials, Client ID Metadata Documents when the authorization server advertises them, or dynamic client registration when configured |
No authentication | Supported ("No sign in") | Supported |
Fixed credentials such as an API key or bearer token | Supported through request headers that Claude sends on every request | Not listed among the documented options, which are OAuth, no authentication, or a mix of the two |
Mixed public and protected tools | Sign-in can be required up front or only when needed | "OAuth or no authentication" lets initialize and tool listing run without auth while individual tools require OAuth |
A few practical consequences follow from this:
OAuth is the most portable choice. If you want one server to work well in both clients, and you want each user to act with their own permissions, support OAuth.
Shared API keys mean shared access. Anthropic points out that on Team and Enterprise plans, a connector using a fixed credential gives everyone who uses it that credential's access. Use a key limited to what members should be able to do.
Your API still enforces permissions. The MCP layer should pass identity through, while your backend keeps checking credentials, authorization rules, and rate limits.
For a deeper look at API keys, bearer tokens, OAuth, and credential pass-through, read API authentication for MCP.
How to connect a remote MCP server to Claude
On individual plans
Go to Customize > Connectors.
Select + Add, then Add custom connector.
Enter a name for the connector and the remote MCP server URL, then continue.
Review the authentication settings Claude detected and adjust them if your server needs something different.
Choose how people connect: Sign in now, Sign in when needed, or No sign in.
If you use OAuth, choose how Claude identifies itself to your authorization server.
If your server uses a fixed credential instead of OAuth, add it under Request headers.
Select Add, then enable the connector in a conversation from the + menu under Connectors.
On Team and Enterprise plans
An Owner or Primary Owner adds the connector once for the organization from Organization settings > Connectors by choosing Add, then Custom, then Web, and completing the same URL and authentication steps. On Enterprise, people with a custom role that includes managing libraries can add one too. Members then find the connector under Customize > Connectors and select Connect to sign in with their own account.
If you change the server URL or authentication later, Anthropic's help center says to remove the custom connector and add it again with the updated details.
Claude Code and developer clients
Developer-facing clients such as Claude Code, Cursor, VS Code, Codex, and Windsurf usually connect through a configuration entry that points at the server URL and supplies any required headers or environment variables. Keep secrets in environment variables or the client's credential store rather than in a shared configuration file.
How to connect a remote MCP server to ChatGPT
Open ChatGPT on the web and go to Plugins.
Select the plus button, then Add custom MCP server.
Enter a name and an optional description.
Under Connection, enter the server URL, or choose Tunnel for a Secure MCP Tunnel.
Configure authentication: OAuth, no authentication, or OAuth or no authentication.
Review the risk warning, confirm, and select Create as a plugin.
Install the resulting plugin from your personal plugins or your workspace.
In a conversation, type @ and select the plugin to use it.
Workspace permissions and security restrictions apply to adding and using custom MCP servers, so on a business workspace an admin setting may decide whether this option is available.
How ChatGPT treats read and write tools
ChatGPT supports both read and write tools. Write actions require confirmation by default, and ChatGPT respects the MCP readOnlyHint tool annotation. Tools without that hint are treated as write actions. If a tool only reads data, mark it read-only so users are not asked to confirm every lookup. If a tool changes data, leave it as a write action and make its description say exactly what it changes.
When you update tools on the server, use the plugin's details page in ChatGPT to refresh it and pull in new tools, descriptions, and server instructions.
Make your server easy for both clients to use
Connecting is only the first step. Whether the model picks the right tool depends on what your server publishes.
Write action-oriented names and descriptions. OpenAI recommends descriptions that include "Use this when..." guidance, note edge cases, and describe parameters with enums where possible.
Use server instructions for cross-tool rules. The MCP
instructionsfield is the place for guidance such as required tool order or shared rate limits. OpenAI advises keeping the first 512 characters self-contained.Annotate read-only tools. This reduces unnecessary confirmation prompts in ChatGPT and makes intent clearer in any client.
Keep results focused. Return the fields a workflow needs instead of a full upstream response.
Version changes deliberately. Users and clients rely on the tools they already discovered, so renaming or removing a tool can break existing workflows. See how API changes affect MCP servers.
Troubleshooting common connection problems
Symptom | Likely cause | What to check |
|---|---|---|
Claude cannot reach the server, but it works from your machine | The server is only reachable on your network | Public reachability, firewall rules, VPN, or Anthropic IP allowlisting |
Connection fails immediately | Wrong URL or unsupported transport | The exact endpoint path and that the server speaks Streamable HTTP |
401 or 403 responses | Missing, invalid, or under-scoped credentials | OAuth configuration, request headers, token scopes, and upstream API permissions |
Connected, but no tools appear | Nothing published, or discovery failed | That the live server version contains the tools, then reconnect or refresh the client |
Tools appear, but the model picks the wrong one | Vague names or overlapping descriptions | Tool names, descriptions, parameter docs, and server instructions |
Old tool list after a server update | The client still has the earlier discovery result | Refresh the plugin in ChatGPT, or disconnect and reconnect in other clients |
Server-side logs help a lot here. If a connection attempt leaves no log entry on your server, the request never arrived, and the problem is in the client configuration, DNS, network path, or URL rather than in your tools.
Where 0mcp fits
If your API is described by an OpenAPI or Swagger file or a Postman collection, 0mcp turns the operations you select into MCP tools and serves them from a hosted Streamable HTTP endpoint, with authentication pass-through so your API keeps enforcing permissions. Its Setup page generates the endpoint and client-specific configuration from the published version, with setup guides for Claude Code, Claude Desktop, Cursor, Codex, VS Code, and Windsurf, and its logs show whether a connection attempt reached the server. You can also read how to deploy an MCP server for production if you are weighing hosting options.
Final checklist before sharing your server URL
The endpoint is public, or you have planned allowlisting or a tunnel.
The server uses Streamable HTTP.
Authentication works in each client you plan to support, and OAuth is in place if you need ChatGPT and per-user permissions.
Only reviewed tools are published, with clear names and descriptions.
Read-only tools carry the
readOnlyHintannotation.You have tested discovery and tool calls outside the client first.
You know how users will pick up tool changes after an update.
FAQ
01Can I connect a local MCP server to Claude.ai or ChatGPT?+-
Not as a custom connector. Claude connects to custom connectors from Anthropic's cloud, so the server must be reachable over the public internet or allowlisted for Anthropic's IP ranges. Local servers configured in Claude Desktop's config file are a separate mechanism. ChatGPT expects a server URL, and also offers a Secure MCP Tunnel option.
02Which transport should a remote MCP server use for Claude and ChatGPT?+-
Use Streamable HTTP. It is the transport the MCP specification defines for remote servers. Anthropic describes HTTP+SSE as legacy and being deprecated, and OpenAI lists both SSE and streaming HTTP as supported for ChatGPT.
03Can I use an API key instead of OAuth?+-
In Claude, yes: add fixed credentials such as an API key or bearer token under Request headers when you add the custom connector. On Team and Enterprise plans everyone using that connector shares the credential's access. ChatGPT's documented options for custom MCP servers are OAuth, no authentication, or OAuth or no authentication.
04Why does ChatGPT ask me to confirm every tool call?+-
ChatGPT treats tools as write actions unless they carry the MCP <code>readOnlyHint</code> annotation, and write actions require confirmation by default. Mark tools that only read data as read-only on the server.
05My server connected but no tools show up. What should I check?+-
Confirm the live server version actually publishes the tools, then reconnect the client. In ChatGPT, refresh the plugin from its details page to pull new tools. If your server logs show no request at all, the problem is in the URL, network path, or client configuration.
06Does 0mcp work with Claude and ChatGPT?+-
0mcp serves hosted MCP endpoints over Streamable HTTP, and its Setup page generates client configuration for Claude Code, Claude Desktop, Cursor, Codex, VS Code, and Windsurf. Any client that supports remote Streamable HTTP servers and the authentication your published tools need can connect to the endpoint.