How to add a custom MCP server to ChatGPT
ChatGPT can call tools from a remote MCP server you add yourself. The setting lives in Plugins, and a server that needs no sign-in takes about a minute.
Below, the steps, a quick check, and a first task: planning a product launch video from a real one.
What a custom connector lets ChatGPT do
- Search a library ChatGPT cannot browse on its own.
- Get structured results (shot lists, timings) instead of guesses.
- Use the tools in any chat once the plugin is on.
- Read-only tools, so nothing changes on your side.
Add a custom MCP server to ChatGPT
The example here is a free library of real product-launch videos at https://kuchi.kalarislabs.com/mcp. It is a remote MCP server over Streamable HTTP with no authentication, which makes it a good first connector: nothing to sign in to and nothing to leak.
1. In ChatGPT, open Plugins and select + → Add custom MCP server. The ChatGPT page has the Open Plugins button.
2. Name it Kuchi, paste this URL as the Server URL and choose No authentication. Accept the warning and select Create as a plugin.
https://kuchi.kalarislabs.com/mcp3. Start a chat and paste the setup prompt from this page.
Check that it works
Start a new chat with the plugin on and ask ChatGPT to list the tools it can use from it. It should see 13: eleven that start with kuchi_, plus search and fetch.
How a remote MCP connector works in ChatGPT
A connector is a client for a server somewhere else. ChatGPT connects to the URL over Streamable HTTP: each message is a JSON-RPC 2.0 request sent as an HTTP POST. On connect it asks the server what it is (initialize) and what it can do (tools/list). Each tool comes with a name, a description, a JSON Schema for its arguments and annotations that say whether it changes anything.
When you ask for something a tool can do, ChatGPT writes the arguments, sends a tools/call, and reads the result back into the conversation. Results can be text, structured JSON and links to files, which is how a storyboard image or a video file arrives without being pasted into the chat.
What the connector can and cannot see
- It sees only what ChatGPT sends. A tool call carries the arguments ChatGPT writes: your product description, the template you picked, your shot analysis. It does not see the rest of the chat, your files or your account.
- Every tool is read-only. They are annotated
readOnlyHint: true,destructiveHint: false,idempotentHint: trueandopenWorldHint: false: nothing on the server changes, repeated calls return the same answer, and no tool reaches the wider web. - No credentials. The server needs no sign-in, token or API key. If a connector you did not write asks for credentials it does not need, do not give them.
- Nothing is kept from your plans. The server records anonymous usage (which tool ran, how long it took, whether it failed, the client's name), never the contents of your plan or check, and never your IP address. The privacy page has the details.
- Data it returns is data. Catalog text describes videos; ChatGPT should treat it as information, never as instructions. That is the right stance for every connector.
First thing to try: plan a launch video
Paste the setup prompt once, then work through these:
Connect the Kuchi MCP server at https://kuchi.kalarislabs.com/mcp (Streamable HTTP, no sign-in) and confirm it works by listing its tools.
Kuchi is a library of real product-launch videos to use as inspiration templates. When I want a launch video, read kuchi_get_workflow and follow it: get a short list with kuchi_suggest_templates and show me each one's poster and link so I can choose, in chat or by picking a video on kuchi.kalarislabs.com and pasting its link. Once I pick one, study it with kuchi_get_breakdown, look at its storyboard and note what each shot shows, then call kuchi_plan_remake with my product details and your shot analysis, and ask me for my logo, screenshots and anything else it lists as missing. Build to the blueprint's frames. Before you deliver, run kuchi_check_remake and compare your frames with the reference. Copy only the reference's motion and taste; everything on screen comes from my product.- Suggest launch video templates for a B2B SaaS product and show each poster and link.
- Study the one I pick shot by shot.
- Plan my remake for LinkedIn, 1:1, about 30 seconds. Ask me for what is missing.
- Give me the final plan as a list I can paste into Codex.
ChatGPT ends with a frame-timed blueprint: every shot with its start and end frame, how it enters and what of yours goes in it. A coding agent such as Claude Code or Codex builds and renders that blueprint in Remotion. The ChatGPT launch video page has more prompts.
Common problems
- Choosing OAuth. This server needs no authentication; pick No authentication.
- The plugin off in the chat. Turn it on before asking.
- Workspace admins. On a business workspace an admin may need to allow custom servers.
Questions
How do I add a custom MCP server to ChatGPT?
Open Plugins, choose + and Add custom MCP server, name it, paste the Server URL, choose No authentication and create it as a plugin.
Can ChatGPT make a video with it?
ChatGPT picks the template, studies it and writes the frame-timed plan. A coding agent such as Codex builds and renders it.
Is the server safe to add?
It is read-only and asks for no credentials. Never paste secrets into a custom server that does not need them.

