Connect Unitcore to ChatGPT

Unitcore connects to ChatGPT the way it connects to any other AI client: you add the server address, sign in to your Unitcore account and approve the request. From then on ChatGPT reaches your data as you, and strictly within the boundaries you see yourself.

There is one difference from Claude worth knowing up front: ChatGPT keeps custom MCP servers behind developer mode, and that has to be switched on first. The rest is the same couple of minutes, once per account.

Connect in one click

This is the only value you will have to type:

No keys, no tokens, no passwords go anywhere — access is granted through your ordinary Unitcore account.

Before you start

• Sign in to Unitcore in this same browser. Then the authorization step takes you straight to the consent screen instead of a login form.
• You need ChatGPT on the web. A custom app can only be created on chatgpt.com. Once created it is available in the mobile app too, but it cannot be added from a phone.
• Any plan will do, including the free one. Developer mode is there on Free as well — it is simply off by default, and nothing points you to it.
• On Business, Enterprise or Edu: a workspace admin enables developer mode first — for themselves, and through role settings for everyone else.

Step 1. Turn on developer mode

Some builds still call the section Connectors rather than Apps — it is the same place.

Step 2. Create the app

Under Apps, press Create and fill in the form.

The address is needed in full, including /api/mcp: there is no MCP server at the domain root.

The name is only a label — it is how you recognize this connection in the list, and it grants nothing. ChatGPT introduces itself to the server as ChatGPT whatever you type here.

Under Authentication, choose OAuth. Below it is a checkbox confirming that you trust this server; you have to tick it. Then press Scan tools: ChatGPT reaches the server, gets a 401 along with the address of the authorization server, registers itself as an OAuth client, and lists the tools it found. Then press Create.

Why the other auth options do not fit

No authentication tells ChatGPT the server answers without a sign-in. It does not: a request without a token gets a 401, so connecting without authorization yields nothing at all.

API key asks for a static key you obtained beforehand. Unitcore has no such path, deliberately — a key living in another service's settings can be neither scoped nor revoked one client at a time.

There is no OAuth client to register ahead of time either: clients register dynamically, and each connection gets its own. You will never see a client ID or a secret.

Step 3. Sign in and approve

ChatGPT opens a Unitcore page — with a login form first, if you are not signed in yet.

Then comes the consent screen. Three rows on it are worth actually reading:

• Client — the name of the client asking for access; here that is ChatGPT;
• Redirects to — where you will be returned; for ChatGPT that is an address like https://chatgpt.com/connector/oauth/…. Every app gets its own, but the domain is always chatgpt.com;
• Signed in as — whether this is your Unitcore account.

Below them is something no client's own dialog offers: what this client may do.

Starting read-only costs you nothing: the level can be changed later in one click, on the connected applications page.

Allow access issues the token; Deny ends the flow and no access is granted. After you approve you are returned to ChatGPT, and the app shows up in the list as a draft — that is expected and does not stop you using it.

Step 4. Turn the app on in a conversation

The app is attached to your account, but each message enables it separately: in the composer press +, choose More → Developer mode and tick Unitcore.

That selection applies to one message. If a follow-up needs fresh data, @-mention the app again. That is ChatGPT's behaviour, not ours.

Step 5. Check that it works

The answer should contain your personal space, created automatically when you signed up. If you see it, you are connected.

Now a write, if you granted one:

Before the call ChatGPT shows you the JSON payload it is about to send and asks you to confirm. That is a built-in developer-mode safeguard, and it is worth reading rather than clicking past: the amount and the category are right there in it.

Why this is safe to connect

Connecting an AI client to your data deserves more than a reassuring sentence. The rule is this: a tool acts as the user who called it, and never as anyone else, and four things keep that rule from eroding:

• Isolation is enforced by Postgres, not by tool code. The request carries your OAuth token, and Row Level Security decides what comes back. Nothing in the tool bodies enforces isolation, so nothing in them can accidentally undo it.
• "Read only" is a database policy too, not a polite request to the model. If a connection is marked read-only, writes fail no matter how insistently the assistant is asked.
• Only OAuth tokens are accepted. Verification requires a client_id claim, so a web session lifted from a browser is not a key to this server. For the same reason write access can only be changed from a browser: a client cannot grant it to itself.
• A service-role key cannot reach the tools. That key bypasses RLS, and the build fails if the MCP code so much as reads an environment variable named SECRET or SERVICE_ROLE.

There is an end-to-end test as well: it drives every read tool as one user and fails if an identifier belonging to another shows up in the answer.

How to disconnect

• Selectively: on Settings → Connected applications in Unitcore, every connection has a Write access toggle. Switch it off and ChatGPT can read and reckon but write nothing.
• Entirely: same page, Revoke access. The refresh token dies at once; an access token already issued keeps working for up to an hour, as the authorization server has no revocation endpoint yet.
• On the ChatGPT side: Settings → Apps → Unitcore → delete. Do both: deleting it in ChatGPT does not revoke the tokens it was issued.

If something goes wrong

Questions people ask

Will the free ChatGPT do? Yes. Developer mode is available on the free plan too — it is off by default, but it is switched on in the same Advanced settings, and a custom app is created after that just as it is on any other plan.

Does it work in Deep Research? Not yet. Deep Research requires two specific tools from the server, search and fetch; Unitcore does not provide them — the connector is built for regular chat.

Does this work on a phone? Yes. ChatGPT reaches Unitcore from its own cloud rather than from your device, so an app created once in the browser works in the mobile app too.

If I connect from a work account, will colleagues see my data? No. Access is bound to your Unitcore account, not to a device or a ChatGPT workspace. Do note that on Business and Enterprise a published app is visible to everyone in the workspace — what they see is the connection form, not your data, and each of them signs in to their own Unitcore account.

Can I connect Claude and ChatGPT at the same time? Yes. They are two separate clients with separate permissions: one can be allowed to write while the other stays read-only. Both appear on the connected applications page.