How to Connect ChatGPT to a Remote MCP Server

ChatGPT and the OpenAI API support connecting to a custom, remote MCP server, but through two distinct surfaces that work differently: the ChatGPT consumer app, and the OpenAI Responses API for developers building their own applications. This guide covers both, and is upfront about which parts of the ChatGPT flow have changed recently, so you check the current UI rather than trusting a screenshot from earlier in the year.

Connecting via the OpenAI Responses API

If you're building your own application and want it to call tools on a remote MCP server, the Responses API has a dedicated mcp tool type. This is the more stable, better-documented of the two surfaces:

{
  "type": "mcp",
  "server_label": "my-server",
  "server_url": "https://your-subdomain.mcpcore.io/mcp",
  "server_description": "My custom MCP server",
  "authorization": "Bearer YOUR_API_KEY",
  "require_approval": "never",
  "allowed_tools": ["get_weather", "search_orders"]
}

Key details:

• Transport: Streamable HTTP or HTTP+SSE are both supported.
• Authentication: the authorization field carries your bearer token or API key on every call. OpenAI does not store it.
• Scope: only the tools capability is exposed through this integration; resources and prompts are not part of it.
• require_approval: controls whether tool calls execute automatically ("never") or need explicit approval, which you can also configure per tool with an object instead of a single string.
• allowed_tools: restricts which of your server's tools this integration can call, useful if your server exposes tools you don't want this particular application to use.
• For servers that aren't publicly reachable, OpenAI documents a Secure MCP Tunnel option for connecting to private or on-premises endpoints.

This is the integration path to use if you're writing code that calls the OpenAI API directly, rather than configuring the ChatGPT consumer app.

Connecting via ChatGPT (Consumer App)

ChatGPT itself supports adding a custom remote MCP server through a Developer Mode setting. This capability is real and documented by OpenAI, but the exact path to it in the Settings menu has changed at least once during 2026 as OpenAI renamed and reorganized this part of the product. Rather than give you a click path that may already be stale by the time you read this, here's what to look for:

1. In ChatGPT's Settings, look for a section related to connectors, apps, or integrations.
2. Within that section, look for a Developer Mode toggle. This is what exposes the option to add a custom MCP server by URL, rather than only OpenAI's built-in connectors.
3. When adding your server, you'll provide the server URL and choose an authentication method. ChatGPT supports no-auth, API key, and OAuth, and the mode you select here must match what your server actually requires. A mismatch here is the most common connection failure.
4. Custom connectors generally need to be enabled per chat session, even after you've added one once. If your tools stop appearing in a new conversation, check whether the connector needs to be turned on again for that session.

On plan availability: access to Developer Mode and custom MCP connectors is gated by ChatGPT plan tier, and which tiers get read-only versus full read-write access has also been reported inconsistently across sources during 2026. Don't assume your plan has access, or that it has full access, without checking ChatGPT's current settings directly.

Given how much of this has moved in the past several months, treat the general capability (Developer Mode exists, it lets you add a remote MCP server by URL with your choice of auth) as solid, and verify the exact menu path and plan requirements against OpenAI's own current documentation before writing internal instructions for your team.

Testing the Connection

Regardless of which surface you're using, confirm your server actually responds before troubleshooting the client side:

curl -X POST https://your-endpoint/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

If this returns a valid tool list but ChatGPT or the API still can't reach it, the issue is almost always the authentication mode mismatch mentioned above, or an HTTPS certificate problem, rather than the server itself.

If You're Hosting the Server

Whichever way you connect, the server side of this is the same MCP server you'd host for any other client. If you're setting one up, our MCP server hosting page covers what a production endpoint needs, and MCPCore provisions the HTTPS endpoint, subdomain, and authentication automatically so you can focus on the ChatGPT side of the setup rather than the server side.

---

OpenAI's ChatGPT settings and plan tiers for custom MCP connectors have changed multiple times in 2026. Always check OpenAI's current Help Center documentation and API documentation rather than relying on a fixed menu path from an older guide, including this one.