Before it acts in one of your tools (sending an email, creating a task, posting a message, deleting a file), the AI can ask for your approval. You set this in the message box, per connector and per agent; some situations require your approval whatever the setting.
Reading happens directly: searching an email, reading a document or checking a calendar, except in the cases where Moon AI always asks.
The message box settingLink to the The message box setting section
In the message box, open “+” › “Connectors”. The “Before acting” section offers two choices:
| Choice | Effect |
|---|---|
| “Approve each action” | Moon AI shows you each email, message, creation or deletion before carrying it out. This is the default setting. |
| “Act without asking” | Connector actions run directly, without your approval. |
When some connectors have their own setting, the section is called “Default permission” and shows how many connectors are exceptions. “Per-connector settings” opens the “Connectors” page.
A connector’s settingLink to the A connector’s setting section
Each connector can have its own setting, which takes priority over the message box setting.
- Open “Connectors” (chat.realmoon.ai/connectors), then the service’s card.
- Under “Permissions”, choose “Use the composer setting”, “Approve each action” or “Act without asking”.
A connector set to “Act without asking” has the “No approval” badge on the “Connectors” page.
Decide in the conversationLink to the Decide in the conversation section
When the AI wants to act, a card appears above the message box and in the answer. It names the action and the service, for example “Send via Gmail?”, and shows the main parameters: recipients, subject, text, file.
| Button | Effect |
|---|---|
| “Approve” | The action runs, then the AI continues its answer. For sending or posting, the button is called “Send”. For a deletion, it shows in red. |
| “Deny” | Nothing runs. The AI is told and carries on without this action. |
| “Don’t ask again for” followed by the service name | The action runs and this connector will act without asking from now on, including to send, create or delete. On a phone, Moon AI asks you to confirm first. |
With an agent, “Don’t ask again for this agent action” replaces the last button: only this action of this agent switches to “Always allow”. You can undo this in the agent’s “Rules” tab or from the “Connectors” page.
To decline an agent’s action, “Deny” opens the “Deny this action?” window. You can write a “Reason (optional)”: the agent reads it and suggests something else. “Always block this action for this agent” creates a “Block” rule.
The “Pending approvals” inboxLink to the The “Pending approvals” inbox section
Requests from agents, routines and messaging apps are not lost if you close the conversation. They wait in “Pending approvals” (chat.realmoon.ai/approvals), in the “Automate” group of the sidebar. The number of pending requests shows next to it, and the notification bell also leads there.
| Item | What it does |
|---|---|
| “To approve” and “Decided” | Pending requests and the history of decisions. |
| “Filter by agent” | Shows only one agent’s requests. |
| Grouped by workspace | Each request stays in the workspace where it was made. |
| “Expires in” | The time left before the request expires. |
| “View conversation” | Opens the conversation where the action was requested. |
A request expires after 72 hours. The “Approval expires soon” notification warns you 24 hours before. An expired request never runs.
The same request can be decided from the conversation, from “Pending approvals” or from a connected messaging app. The first decision counts: the others show “This action was already decided or has expired.”
An agent’s rulesLink to the An agent’s rules section
An agent can have its own rules per connector and per action: “Default”, “Always allow”, “Ask”, “Block”. You set them in the “Rules” tab of the agent’s page. See Agents.
When Moon AI always asksLink to the When Moon AI always asks section
In these cases, Moon AI asks for your approval even if the connector or the message box is set to “Act without asking”, and “Don’t ask again” is not offered.
| Situation | Why |
|---|---|
| A community agent wants to act | Its instructions were written by another user. |
| You use an agent shared by a colleague | The agent is not yours: each action waits for your approval. |
| The agent relies on files or a skill written by someone else | This content may contain instructions you did not write. |
| An email you received was read in the conversation | An email comes from a third party: every action that follows needs your approval, until the end of the conversation. Tools from your MCP servers also ask for your approval, even read-only ones. |
| The conversation contains content received through a webhook, a messaging app or a trigger | This content comes from outside. |
| The conversation contains a deep research report | The report draws on web pages. |
| An agent, or Agent mode, read a web page or a web search result in the conversation | A web page could dictate the action: every send, creation or deletion that follows needs your approval, whatever the agent’s rules. |
| The action has the “Approval required” badge | Sensitive actions: money (refund, payment, invoice, subscription), bulk sends, public links and file sharing, permanent deletions, discount codes, DNS records. |
| Agathe, the appointment scheduling agent | Every appointment she creates, moves or cancels. |
| An Archicad command that changes the project | Every command. See Design software. |
| A send to other people (email, message, post, share) right after reading outside content in the same answer | A web page, a search, a message that was read or an attached file could dictate what gets sent. |
Actions marked “Approval required” are flagged on the connector’s card, under “Show tools”.
When an action is blocked without asking youLink to the When an action is blocked without asking you section
In some cases, the action is not even offered for your approval:
- a routine without the “Allow actions” option cannot send, create or delete anything;
- a routine started by a webhook does not send, create or delete anything in your tools, whatever the option;
- during an agent task trial, actions are simulated: nothing goes out;
- an action set to “Block” in the agent’s rules never runs: the agent looks for another way.
In each case, the AI is told that the action was not carried out.
From a messaging appLink to the From a messaging app section
When you talk to an agent from Telegram, each action it prepares arrives in Telegram with the “Approve” and “Decline” buttons, and in “Pending approvals”. See Messaging apps and triggers.