How do I add an MCP connector to ChatGPT?

Turn on Developer mode in Settings → Security and login, open the apps/connectors page, create a new developer-mode app, and paste your MCP server's URL. Complete any sign-in the server asks for, then select the app from the Developer mode option in the composer when you want to use it.

Why can't I see Developer mode in my ChatGPT settings?

Three usual causes. You are on the mobile app — Developer mode is web only. You are on a plan without it: OpenAI lists Pro, Plus, Business, Enterprise and Education. Or you are in a workspace where an administrator has disabled it.

What URL do I paste for an MCP server?

The server's remote endpoint, which usually looks like https://example.com/mcp or https://example.com/api/mcp. It must be a remote server over SSE or streaming HTTP — a local stdio server that runs on your own machine cannot be reached by ChatGPT on the web.

ChatGPT is not calling my MCP tools. What's wrong?

Most often the app is added but not selected for the conversation, or the model is choosing a built-in tool instead. Select the app from Developer mode in the composer, then be explicit in your prompt: name the app and the tool, and say not to use other tools.

Do I need to write code to add an MCP connector?

No. Adding an existing MCP server is entirely a settings task — paste a URL and sign in. Writing your own MCP server is a development project, but connecting to someone else's is not.

How do I remove an MCP connector from ChatGPT?

Open the app's details page in settings and delete it. That removes ChatGPT's access. If the server used OAuth, revoke the connection on the server's own side too — deleting the app in ChatGPT does not always revoke the token the server issued.

A how-to · ~6 min read

How to add an MCP connector to ChatGPT

It is a settings task, not a coding task. Turn on Developer mode, paste a URL, sign in. Five minutes. This walkthrough works for any remote MCP server — the Konshus URL is used as the worked example.

Before you start

Use the web app. Developer mode does not exist in ChatGPT on mobile.
Check your plan. OpenAI lists Pro, Plus, Business, Enterprise and Education.
Have the server URL ready. It must be a remote endpoint over SSE or streaming HTTP. A local stdio server cannot be reached from the web app.

The steps

Turn on Developer mode. Settings → Security and login → Developer mode. Read the warning — OpenAI classifies this as elevated risk, and that is not boilerplate.
Go back to the chat screen and open Plugins. It's in the left sidebar (under More if you don't see it). The plus button at the top right of the Plugins page only offers developer-mode apps once the toggle above is on.
Create the app and paste the URL. For Konshus that is https://konshus.ai/api/mcp. ChatGPT discovers the OAuth settings itself — leave them as found. Notices about CIMD or OIDC being unavailable are normal for servers that use Dynamic Client Registration.
Authenticate. If the server uses OAuth you will be sent to sign in and approve access. Servers with no authentication skip this — which is also a reason to be careful about which ones you trust.
Check the tool list. The app's details page shows every tool the server exposes. It may say "No app actions available yet" for a minute right after connecting — wait, then use Refresh. Use refresh here too after the server ships new tools.
Use it. In a conversation, choose Developer mode from the composer menu and select the app.

One more step: set your permissions in ChatGPT

Give it a minute. After you approve, ChatGPT may show "This connector has no tools available" for up to a minute while it fetches the tool list. That's normal and it's on ChatGPT's side — don't hit Disconnect. Close and reopen the connector and all tools will be there.

Just added a new capability? Refresh the connector.

ChatGPT reads the list of things your Vault can do once, when you first connect it, and then remembers it. If ChatGPT tells you a Konshus capability "isn't available", it's working from the old list.

Fix it inside ChatGPT (not here in Konshus): ChatGPT → Settings → Apps & Connectors → Konshus → Refresh. Give it a minute, then start a new chat.

New: ChatGPT can now read a whole document in your Vault end to end — a filing, a contract, a long export — not just find it and quote a snippet. If it says that capability isn't available, refresh the connector above. Anything you've marked Private stays unreadable, and ChatGPT will say so by name.

In ChatGPT this is one setting, not a list of tools. Go to Settings → Plugins → Konshus → Permissions and pick how often it should stop and ask you.

Allow low-risk actions (the default)

Reads mostly go through, but ChatGPT can still stop and ask before saving anything back to your Vault. Fine to start here.

Allow read actions — the minimum we'd pick

Your Whisper loads silently at the start of a chat instead of interrupting you with a popup. ChatGPT still asks before anything is written.

Allow all actions — on the AI you use every day

This is what makes your Konshus grow without you babysitting it: things worth remembering get saved back as you talk. ChatGPT tags it elevated risk — that's a blanket label it puts on every connector, not a judgement about Konshus. Nothing we expose deletes anything, and everything that lands is reviewable in Sources. Leave it on "Allow read actions" if this is a shared or second machine.

Right after you connect, this page may say "No app actions available yet". That's ChatGPT still fetching the tool list — wait a minute and hit Refresh.

When ChatGPT ignores your tools

This is the most common complaint and it is almost never a broken connection. The model has many tools available and picks what it thinks fits. Be explicit:

Use the "Konshus" app's search_atoms tool to find what you know about my writing style. Do not use browsing or any other tools.

Naming the app, naming the tool, and explicitly ruling out alternatives fixes the large majority of "it did not call my server" reports. OpenAI recommends the same pattern in their own guidance.

Write actions deserve a pause

Read tools are low stakes — worst case you get an unhelpful answer. Write tools can delete, overwrite, or share real data, and the model can get the arguments wrong. ChatGPT asks for confirmation on writes by default. Leave that on, and actually read the payload it shows you before approving. Expanding the tool call to inspect the JSON takes two seconds and has saved people real damage.

Frequently Asked Questions

Need a server worth connecting?

Konshus is a memory vault you own, served to ChatGPT over MCP. Paste one URL and the model stops guessing at who you are.

Meet Konshus