> ## Documentation Index
> Fetch the complete documentation index at: https://docs.yorlet.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Agents and Inbox

> How the assistant, email agent, Recover triage, Claude, and ChatGPT use Inbox threads.

Agents do not have a separate mailbox. They read and write the same [threads](/business-automation/inbox/overview) you see in Inbox. On the email channel they send from a [provisioned address](/business-automation/inbox/email) on your account.

Set the address they send from in [Settings → Agents](https://dashboard.yorlet.com/settings/agents), under **Agent email**. You can change it at any time. If you delete that address, pick another provisioned one before agents send again.

<Note>
  The first address you provision is marked as the agent address automatically. Later addresses stay available for people to send from; they are not the agent address unless you select them here.
</Note>

## How agents send

Every agent send on email — from the [assistant](/business-automation/ai/assistant), the email agent, [Recover triage](/billing/recovery), or [Claude or ChatGPT](/business-automation/ai/mcp) — creates a **Send email** [approval](/business-automation/ai/approvals) when [policy](/business-automation/ai/policy) is **Require approval**. That is the default for sending on the email channel.

The approval shows the **To**, **From**, **Subject**, and body. Approving it sends the message and writes it to the thread in [Inbox](/business-automation/inbox/overview). Denying it leaves the thread as it was.

People sending from the Dashboard or [API](/development/api-keys) with the **Send email** permission do not go through Approvals. That path is a human send, not an agent action.

## Email agent

When someone emails the **agent address**, Yorlet:

1. Matches the sender to a customer or owner when it can.
2. Adds the message to a [thread](/business-automation/inbox/overview) in Inbox.
3. Skips automatic mail such as out-of-office replies.
4. Drafts a reply and creates a **Send email** approval.

Approve the draft from [Approvals](/business-automation/ai/approvals) or from the card in the assistant. If the situation needs a person — hardship, a complaint, or missing context — the agent proposes an internal note instead of a reply.

Mail to addresses that are not the agent address is still stored in Inbox. The agent does not draft a reply for those.

<Tip>
  Point customers at the agent address when you want inbound mail to produce a draft on that thread.
</Tip>

<Note>
  Yorlet will not keep auto-replying in a loop. Automated mail is ignored, and a thread has a daily cap on outbound messages.
</Note>

## Recover triage

[Recover](/billing/recovery) can read a customer's existing Inbox threads before it proposes the next step. If a conversation is already underway, it can propose a **Send email** on that thread instead of opening a new recovery case. Those sends use the same agent address and the same Approvals queue.

## Assistant, Claude, and ChatGPT

The [assistant](/business-automation/ai/assistant) can send on a thread when you ask it to, using the agent address and optionally an existing conversation in Inbox.

[Claude and ChatGPT](/business-automation/ai/mcp) that connect over MCP only see write tools — including **Send email**. Calling that tool still creates an approval. Give them an [agent key](/business-automation/ai/agent-keys), not a secret API key. See [MCP](/business-automation/ai/mcp) for how to add Yorlet as a connector.

<Warning>
  Do not switch **Send email** to **Auto-execute** until you have watched a run of drafts. The default **Require approval** mode is there so a bad recipient or body cannot leave the account without someone looking at it.
</Warning>