How to add a custom MCP server to ChatGPT
Connect Tiny Plates to the ChatGPT app in four steps, with screenshots.
By Chris Artmann ·

ChatGPT is good at meal ideas. But ask it for a week of dinners and the recipes come from memory. The amounts are guesses, and the shopping list is added up by a language model.
An MCP server fixes that. It gives ChatGPT tools it can call while it works. Add the Tiny Plates server and ChatGPT searches real recipes, scales them for you and builds shopping lists from the actual amounts.
This guide walks through adding it in the ChatGPT app, step by step. It takes about two minutes. The same steps work for any remote MCP server that signs in with OAuth. Only the name and the URL change.
What you need
- The ChatGPT app.
- A Tiny Plates account. You can create one for free when ChatGPT asks you to sign in.
- The server URL:
https://api.tinyplates.dev/mcp
1. Open Plugins in Settings
Open Settings and choose Plugins in the sidebar, under Integrations. This is where ChatGPT keeps your apps, skills and MCP servers. Choose Add at the top right, then Add MCP server.

2. Enter the name, type and URL
Fill in the form:
- Set the name to Tiny Plates.
- Pick Streamable HTTP as the type. The other option, STDIO, is for servers that run on your own computer.
- Paste
[https://api.tinyplates.dev/mcp](https://api.tinyplates.dev/mcp)as the URL.
Leave the bearer token and the headers empty. You sign in with your account in a moment instead. Then save the server.

3. Authenticate
Open the MCPs tab. Your new server is in the list as tiny_plates. ChatGPT turns the name into an id, so don't worry that it looks different. Choose Authenticate next to it.

4. Sign in with Tiny Plates
ChatGPT opens Tiny Plates in your browser. Sign in, or create an account, and allow access. You're sent back to ChatGPT, and the server is ready. There's no API key to copy.
Try it
Start a new chat and ask for something that needs real recipes:
Plan four dinners for this week. Two of us, no milk or peanuts, nothing over 40 minutes. Then give me one shopping list.
ChatGPT searches with your allergens filtered out, checks the cooking times, scales each recipe to two people and asks the API for one combined shopping list. The amounts on the list are real, not estimated.
A few more to try:
- "I have chicken, rice and a lime. What can I make?"
- "Halve this recipe and switch it to metric."
- "I'm out of buttermilk. What can I use in these pancakes?"
If something goes wrong
- If ChatGPT can't reach the server, check the URL. It ends in
/mcp, and it'sapi.tinyplates.dev, notwww. Also check that the type is Streamable HTTP. - If the browser didn't open, go back to the MCPs tab and choose Authenticate again.
- If you can't sign in from this machine, use an API key instead. Under Headers, add
Authorizationwith the valueBearerfollowed by your key, and skip Authenticate.
Every tool call is one API request. It counts toward the same free allowance as the API and shows up on your usage page.
OpenAI moves these settings around now and then. If yours look different, OpenAI's guide to custom MCP servers has the current path.
Other apps
Using Claude instead? See How to add a custom MCP connector to Claude. Claude Code, Cursor, Codex, VS Code and more are in the MCP docs.