🔌 Plugging an MCP Server into the AI Assistant

An MCP connector gives the AI assistant extra tools: reading your Notion pages, querying a business service, triggering an action somewhere else. Once plugged in, the assistant uses it on its own in the Text chat.

🧭 Where This Happens

Open the AI Generation window (shortcut Cmd+Shift+G), then the Tools tab: your connectors are shown there as cards, with the Add a connector button.

➕ Creating a Connector

The button opens the Custom MCP connector dialog, with four fields:

• Title: the name shown on the card. Required.
• Server: the address of the MCP server. Required.
• Description: optional.
• Authentication method: see below.

The Connect button creates the connector, then Celestory queries the server to discover its tools. The connector sheet then opens, with two tabs: Overview and Functions.

🔐 The Four Authentication Methods

• No authentication: nothing to fill in.
• API token: you paste a token, presented to the server on every call.
• JSON headers: your own headers, as a JSON object. Malformed JSON is flagged before saving.
• OAuth: no secret to type in. Celestory queries the server, registers itself with it and takes you to its consent page.

When you pick OAuth, the dialog displays: "The authorization opens in a window right after the connector is created." You approve on the service's own site, the window closes, the sheet switches to Connected and two buttons appear: Authorize again and Disconnect.

🔺 If your browser blocks pop-up windows, nothing starts: allow pop-ups for the site, then try again.

🔺 An authorization left pending for more than ten minutes expires, and has to be started over from Celestory.

Your secrets stay server-side. A stored token is never sent back to the interface: the fields come back empty, and only the method used and the connection state are displayed. Typing a value again replaces the old one; typing nothing leaves it in place.

Renewal is automatic. Celestory renews access before it expires, and if the service still refuses a call, it retries once with fresh access. You only step in if the service granted no renewal at all: the connector stops answering, and Authorize again restores access.

🧰 The Functions Tab

It lists the tools exposed by the server, with their description and a counter. Refresh tools runs discovery again.

Each tool carries an Allow switch. Turning it off leaves the tool visible in the list, but it is never offered to the assistant again: handy for letting nothing but reads through.

🎚️ Switching Off Without Deleting

In the Overview tab, the Enabled switch turns off the whole connector: its card goes to Disabled and the assistant ignores it, without losing its configuration or its authorization. Delete asks for a confirmation.

👥 Connectors Belong to the Project

A connector is attached to the project, not to your account. Anyone who can edit the project sees the same connectors, and an authorization granted by one member serves the whole team. Conversely, a connector does not show up in your other projects.

🔺 Not Every Model Can Use Tools

Some models cannot call tools. If you select one while a connector is active, a banner appears in the Tools tab and in the Text chat. The model will answer normally, but will use no connector at all. Pick a tool-capable model.

📏 Limits Worth Knowing

• The number of connectors is not capped.
• At most 128 tools are presented to the assistant for a given turn; beyond that, the rest are ignored.
• The assistant chains at most 8 rounds of tool calls for a single answer.

🛠️ When It Does Not Work

If the server is unreachable or refuses the connection, the card switches to Connection error and the exact message is shown in the sheet. During a conversation, a failing connector is simply set aside: the assistant carries on with the others.

→ Next step: in the Functions tab, switch off the tools you do not need, then ask the assistant for something in the Text chat to watch it use them.

Updated on: 22/09/2026