Custom connectors: add your own MCP server to ChatGPT and Claude

Before you start

A custom connector is an MCP server you add yourself, instead of choosing an app from a list. MCP, the Model Context Protocol, is the open standard that lets an AI app call another program's tools. A remote server lives at a web address. A local one runs on your own computer, and the MCP course covers that case and the Claude fields in detail.

You need the server's address, a public HTTPS URL that usually ends in /mcp, and the way it signs you in. Mindket lists many servers on its MCP servers page. To write a server yourself, start with the Python lesson.

OpenAI adds one condition: account and workspace policies apply to adding and using custom MCP servers. At work, your administrator may have switched this off. Ask before you try.

Add a custom MCP server in ChatGPT

These steps follow OpenAI's page on connecting and testing a plugin. As that page says, account and workspace policies apply.

1. Open Plugins and press the plus button. OpenAI names chatgpt.com/plugins as the place.
2. Choose Add custom MCP server.
3. Give it a name and a description. OpenAI calls them user-facing, so write what you would want to read in the list.
4. Under Connection, enter the address. Choose the public endpoint and type the server URL, including the /mcp path. For a server that is not public, choose Tunnel and pick a tunnel or enter its tunnel_id.
5. Set authentication. Use whatever the server's owner says it needs.
6. Read the risk warning. Select "I understand and want to continue" only if you trust the server.
7. Select Create as a plugin, then review. ChatGPT shows the tools and metadata it found on the server. Read the tool names before you use any.

Test it with five kinds of request

OpenAI's page says to install the new plugin, start a new conversation, type @ and pick it, then run a test set. Keep the list short. The set covers direct requests, indirect requests, follow-ups, write actions that need confirmation and unsupported requests. For each, write down the tool ChatGPT chose, its arguments, the result, any errors and whether it asked for confirmation.

The kinds of request are OpenAI's. The examples in the table are Mindket's own and assume a calendar server. Swap in your server's job.

The same server in Claude, and one that will not fit

Claude takes a remote server as a custom connector. In Claude Desktop the menu path is Settings, then Connectors, then Add custom connector, as GitHub's install guide gives it. The connect lesson walks through adding a custom connector in Claude by name and URL, and notes that free plans are limited to one. Anthropic keeps a GitHub repository where it posts changes to how Claude handles MCP, such as SSE deprecation and Streamable HTTP updates.

Not every remote address works as a connector. GitHub's own remote server, at https://api.githubcopilot.com/mcp, needs OAuth through a registered GitHub App or OAuth App. GitHub's Claude Desktop note calls that sign-in "not currently supported", says it is investigating and points to the local Docker setup instead.

Anthropic's engineering blog treats anything outside its directory as untrusted. Its advice is to run such a connector against fake data first.

A server on your own computer: test it, then give it an address

OpenAI suggests testing a server with MCP Inspector before you connect it. Its Apps SDK examples show how to expose a local server on port 8000: run ngrok and you get a public address of the shape https://example-name.ngrok-free.app/mcp. The README says to add it under Settings, then Connectors, after enabling developer mode. Test first, then expose.

Python servers add one step. The Python MCP SDK has DNS rebinding protection, so the README says to set MCP_ALLOWED_HOSTS to your tunnel host and MCP_ALLOWED_ORIGINS to its origin before the server starts. The MCP Inspector lesson shows how to list a server's tools and what to check when one is missing from the list.

npx @modelcontextprotocol/inspector@latest
ngrok http 8000

What developers call it: the API's remote MCP tool

Developers meet the same idea in OpenAI's Responses API, and its parameter names show what you control.

• server_url is the address of a remote MCP server that supports Streamable HTTP or HTTP/SSE. OpenAI's examples add a server_label next to it.
• allowed_tools limits which of the server's tools are imported, and OpenAI says that also cuts cost and latency.
• require_approval is always, never or a list of tools that skip approval. By default OpenAI asks for approval before any data is shared with a connector or remote server.
• authorization carries an OAuth token for servers that need one, and the API does not store it.

What changed lately for people who add their own server

Three entries in OpenAI's changelog matter to anyone who adds a server of their own. The May entry says ChatGPT reads the instructions a server returns when it starts and uses them with the tool metadata, so a server's owner can state its constraints once. The August entry gives apps a stable OAuth callback, which concerns whoever builds the server's sign-in. If you only connect servers that someone else runs, you can skim the dates.

Common mistakes

• Leaving off /mcp. OpenAI says to enter the URL with the path, and Cloudflare's guide uses /mcp in its examples.
• Opening the /mcp address in a browser to test it. Cloudflare says you cannot interact with the server that way, so use the Inspector.
• Typing a localhost address. ChatGPT needs a public HTTPS endpoint or a tunnel, as OpenAI's page says.
• Changing the server's tools and not pressing Refresh. OpenAI says to refresh the connection and start a new conversation, and that this applies only to servers you connected directly.
• Adding a server because a page asked you to. OpenAI says servers it has not verified can access, send and receive data and take action.

Try it yourself

• Run the Inspector against a server you trust and read its tool list before you add it anywhere.
• Add that server to ChatGPT, run the five kinds of request and note which one fails.
• Add the same address to Claude and compare the tool list with what ChatGPT showed.
• Open the server's documentation and write down who runs it and where it stores what you send.

Next: what to check before you connect

A custom connector is a server you chose to trust. The last lesson lists what to check first, and shows how a hidden instruction in an email or a document can steer an AI that has a connector.

Frequently asked questions

Sources

1. OpenAI: Connect and test your plugin (read November 6, 2026)https://developers.openai.com/plugins/deploy/connect-chatgpt.md
2. GitHub: openai/openai-apps-sdk-examples (read November 6, 2026)https://github.com/openai/openai-apps-sdk-examples
3. OpenAI: Tools, connectors and MCP (API guide) (read November 6, 2026)https://developers.openai.com/api/docs/guides/tools-connectors-mcp
4. OpenAI: Plugins security and privacy (read November 6, 2026)https://developers.openai.com/plugins/guides/security-privacy.md
5. OpenAI: Plugins changelog (read November 6, 2026)https://developers.openai.com/plugins/changelog
6. Cloudflare: Build a remote MCP server (read November 6, 2026)https://developers.cloudflare.com/agents/guides/remote-mcp-server/
7. GitHub: Install the GitHub MCP server in Claude (read November 6, 2026)https://github.com/github/github-mcp-server/blob/main/docs/installation-guides/install-claude.md
8. GitHub: anthropics/claude-ai-mcp (read November 6, 2026)https://github.com/anthropics/claude-ai-mcp
9. Anthropic: How we contain Claude (read November 6, 2026)https://www.anthropic.com/engineering/how-we-contain-claude