Inbox agent

When someone replies, the inbox agent either drafts a reply for you to review or, on threads your campaigns started, answers from your sales playbook on its own after a delay you can cancel.

The inbox agent answers the people who reply to your outreach. It runs in one of three modes:

Off. Replies are left to you.
Review drafts. On every human reply, the agent drafts a reply in your voice and attaches it to the conversation. You approve and send it, edit it first, or discard it. Nothing goes out unless a person approves it.
Auto-send. On threads your own campaigns started, the agent writes a reply grounded in your sales playbook, checks it, and sends it after a short delay you can cancel. Opt-outs and removal requests are handled before any reply is written. Anything it cannot answer safely falls back to a draft for review.

This is a paid feature and it is off until you turn it on.

Choosing a mode

Open Settings, Workspace and find Inbox agent (it needs the Manage settings permission). Pick a mode under When someone replies, and choose the Playbook the agent works from. Review drafts work without a playbook; auto-send needs one.

Paid and opt-in

The inbox agent is available on paid plans and is off by default. Each reply it writes costs 5 AI credits; opt-outs and removal requests cost nothing. See AI credits.

What the agent answers

The agent only ever answers someone who already replied to you. It never starts a conversation and never writes to a new address. It is asked about a reply when:

a person replied to one of your campaign emails (out-of-office notices, auto-replies and bounces are filtered out before the agent sees them, and are never answered)
your workspace is on a paid plan and the agent is not off
you have at least 5 AI credits, for a reply that needs writing
the reply has enough substance to answer: a one-line acknowledgement ("thanks", "got it") is skipped, so you are never charged to answer a throwaway line

In review drafts mode, a reply the tagging classifier reads as closed, a legal threat, or one that needs a person's judgement is not drafted at all.

Auto-send

Which threads are answered alone

Auto-send replies on its own only on a thread your campaign started: the reply has to answer an email one of your campaigns sent to that person, or a reply the agent itself sent them. Everything else (a fresh email from a contact, a reply matched to a campaign only by the contact's address, a reply from someone copied on the thread) gets a draft for review instead, even in auto-send mode.

What happens to each reply

The reply	What the agent does	Credits
Opens with "stop", "unsubscribe", "no thanks", "not interested", "remove me", "take me off" or "don't email me", uses any reply opt-out phrase, or the classifier reads it as an unsubscribe	Adds the address (and the lead's address, when they answered from another one) to the suppression list, marks the contact unsubscribed, sends nothing	0
Threatens legal action, or asks you to delete their data	Suppresses the address and queues your removal note, on a thread your campaign started, once per thread: a later message (an auto-acknowledgement echoing the request, say) gets nothing more. Anywhere else nothing is sent and a person should read it	0
Is closed or a one-line acknowledgement	Nothing	0
Is on a thread your campaign did not start, or needs a person's judgement	Drafts a reply for review	5
Anything else on your own thread	Writes a grounded reply, checks it and queues it to send	5

Hard stops never wait on the model: an opt-out is honoured even when you are out of credits.

How a reply is written

The agent reads the whole thread, what Zexr knows about the contact, your workspace voice profile, and the full text of your playbook. It then lists every question in the prospect's last message and answers each one from a named playbook section, from the facts it was given (who you are, what you sell, who the contact is), or says plainly that it cannot confirm that detail. It never invents a price, a feature or a promise. The prospect's words are treated as data, never as instructions, so a message that tries to talk the agent into something else is just answered as an email.

Every reply is written in the prospect's language, as plain text, and signed with the sender's first name only.

The checks every reply passes

A reply is only queued when all of these hold:

every answer cites a playbook section that exists, FACTS, or NONE, and a reply to a question cites at least one
every link points to a host on your allowed links list (subdomains included), counting any bare host (example.com, example.top/page, written with any kind of dot) and a mailto: address as a link, though a plain email address is not; with an empty list, replies carry no links
it is between 20 and 6,000 characters
no placeholder is left in it ({{...}}, [NAME], TODO)
it does not mention a playbook, a script, instructions or being an AI
it promises no later follow-up from a person ("get back to you", "check with the team", "circle back", "let me check", "someone will reach out")
the last line is exactly the sender's first name
it has no em or en dash

When a check fails, the agent is asked once more and told exactly what to fix. If the second attempt fails too, your fallback reply is sent instead, after passing the same checks. If even that cannot pass, the reply waits for review with the reasons listed on it.

The send delay and cancelling

A queued reply waits for the send delay (2 minutes by default, 30 seconds to 24 hours) before it goes out. While it waits, the conversation shows the reply, a countdown and Cancel and review: cancelling stops the send and turns it into an ordinary draft you can edit, approve or discard.

Just before sending, the agent checks again and hands the reply to a person instead of sending when:

auto-send was turned off or switched to review drafts
the workspace's plan no longer includes the agent
the conversation moved on: a newer message arrived, or someone queued a reply on the thread by hand
the thread already had as many automatic replies as replies per thread allows (3 by default)
the workspace reached its daily limit of automatic replies in the last 24 hours (200 by default). Each send takes its place in the limit before it goes out, so several servers sending at once never exceed it
the send itself was refused, for example by the suppression list or the mailbox's sending limits

If a send was stored but could not be confirmed, the draft says so and cannot be approved, because the reply may already be on its way: check the thread and discard it once you see it went out. A send that never started is never retried automatically either; it waits for review.

The reason is shown on the draft. A newer reply from the prospect also replaces a reply still waiting to send, so they are always answered on their latest message.

The reply is sent as the owner of the mailbox the prospect wrote to, through the same path as a reply you send yourself, so the mailbox's sending limits, the suppression list and your undo window all apply.

Templates

Fallback reply. Sent when the agent cannot write a reply that passes its checks twice. The default asks for a short call.
Removal note. The one message sent to someone who asks to be removed or threatens legal action. The default reads: "Understood, and sorry for the trouble. I have taken you off our list, and you will not hear from us again." followed by the sender's name. It is the only email that may reach an address after it was suppressed, and only that one person.

Both take two merge fields: {{first_name}} (the prospect's first name, or "there") and {{sender_name}} (the sender's first name), which must be the last line. A template is checked against the same rules as a written reply when you save it.

Who signs

Replies are signed with the first name of Sign as, or, when that is empty, of the sending mailbox's display name. A mailbox with no name sends nothing on its own; its replies wait for review.

Reviewing a draft

A draft waiting for review shows up two ways:

In the conversation itself: open the thread in the unibox and the draft appears below the messages, in an editable box, with Approve and send and Discard. A draft that auto-send held back says why above the text.
In the scope rail: the Agent drafts filter lists every conversation with a draft waiting, with a live count next to it.

From the draft box you can:

Edit the text right there before sending.
Approve and send. Sends the text as it stands (a removal note is sent as written), threaded to their reply, from the mailbox they wrote to (a reply in a shared reply inbox is answered from the mailbox that emailed them), and as that mailbox's owner, whoever approves it.
Discard. Dismisses the draft without sending.

Each thread holds at most one draft in play (waiting for review, queued or sending) at a time. After the agent sends a reply on its own, the conversation shows Sent by the inbox agent with the time, and the workspace audit log records every reply it queued, sent, held back or declined.

What the draft is grounded in

The agent writes with the same context as an AI reply draft: the full thread so far, what Zexr knows about the contact, your workspace voice profile, and your playbook. Without a playbook, review drafts use your enabled skills by name. The more you fill in your voice profile and playbook, the more the replies sound like you and answer what was asked.

Credits and failures

5 credits per written reply, charged when the reply or draft is stored. The retry after a failed check is included. If the model is unavailable, nothing is stored and nothing is charged.
Free hard stops. Opt-outs and removal requests make no model call and cost nothing.
Out of credits. When your balance is below 5, the agent does not write that reply. Opt-outs are still honoured.

Turning it off

Set When someone replies to Off. New replies stop getting drafts immediately. A reply already queued to send is handed to you for review instead of being sent, and drafts already waiting stay until you approve or discard them.