​
Reply

1

Pick a conversation

Click a row in the left panel.
2

Choose the account you send from

The picker above the composer lists your credentials, narrowed to channels this customer actually has a contact on. It defaults to Previous account — whichever account last sent or received on this conversation — so a reply leaves the same mailbox the customer has been talking to.
3

Write and send

Type in the composer, then press Cmd/Ctrl + Enter or click send.
The account you pick decides the channel. Choose an SMTP credential and the reply goes out as email; choose the LinkedIn credential and it goes out as a DM.
If the picker is empty, the customer has no contact on any channel you hold a credential for. Add a contact in the customer drawer, or connect the matching credential under Integrations.

​
AI drafts

An inbox conversation with the AI draft card marked 1, its send button marked 2, the sending account marked 3 and the auto-reply switch marked 4
1

The draft

When the agent is not confident enough to send on its own, it leaves an amber card reading AI draft · needs review, with the confidence it gave itself and a line on why it hesitated. The text is editable in place.
2

Send or discard

Send sends it, Discard throws it away. Auto-reply stays off for that customer until you do one or the other.
3

The account it leaves from

The picker above the composer decides which of your accounts sends, and so which channel the reply uses.
4

Auto-reply for this person

Auto-reply on toggles the agent for this customer alone. It is the switch to reach for when a conversation turns into a sales call.
When the agent hands off without writing anything at all, the card says so instead and no draft text appears.

​
What the AI did

Every tool the agent called lands in the timeline as one card: AI worked here · N steps, with a one-line summary. Expand it to see each step, what it was given, and whether it failed. While the agent is running, the conversation row and the bottom of the timeline name the tool it is on right now, such as AI · Searching knowledge base…, and the list header counts how many conversations it is working on. Reloading mid-run keeps that state — the trace is stored on the server. Hide the whole trace with the ✨ button in the conversation header. That choice is remembered in your browser.

​
Email rendering

Email bodies are rendered as email, not as raw markup:
  • HTML is sanitized and rendered inside a sandboxed frame with no scripts.
  • Raw MIME from the provider — headers, boundaries, base64 and quoted-printable — is decoded, and the HTML part is preferred over the plain-text part.
  • Quoted reply history is folded behind a ••• toggle, and a long email behind Show full email.
  • Remote images stay blocked until you click Show blocked images, so tracking pixels do not fire when you open a message.

​
Scrolling

The timeline follows new activity only when you are already at the bottom. Read further up and nothing moves under you; a New messages pill appears instead, reading Jump to latest when you scrolled up yourself. Email frames that grow as their images load do not shift your position either.

​
Per-customer settings

The ⚙ icon on the contact card opens a drawer holding a short internal note about the customer, up to 500 characters. The note is the same field the Human Review modal appends to, and the AI agent reads it as context.
The auto-reply switch and the tag input inside this drawer do not save yet. Use the Auto-reply on / off button in the conversation header to control the agent, and the Tags section of the customer drawer for tags.