An inbox your AI agent answers, within your rules.

• Someone emails your agent at its address on your domain.
• The agent prepares a reply: it answers the question, confirms the booking, or takes the next step through the tools you connect.
• Both directions are screened: incoming mail for phishing and manipulation, outgoing replies for leaks and policy breaks.
• You decide what it may send: drafts only, approval on every reply, or routine replies on their own.

Each mailbox has its own instructions (your services, hours, FAQs and booking rules), rules for what the agent must never do, and a signature. Replies say they come from an AI assistant, and stay in the thread the customer started.

• Incoming: every message is checked for phishing and for instructions planted for the agent. Phishing is quarantined, and the agent never sees it.
• Outgoing: every reply is checked for data leaks, policy violations and signs the agent followed planted instructions, before anything is sent.

• Drafts only: writes drafts. Never sends or runs tools that need approval.
• Approve everything: every reply and approval-only tool waits for a person.
• Guarded: sends when every check passes; anything risky waits. The default.
• Autonomous: sends unless a hard check fails, such as a planted instruction, a data leak or paused sending.

Held replies wait in a queue in the dashboard, where anyone on the team can approve, edit or reject them.

Connect your booking system, CRM or anything with an MCP server, and choose per mailbox which tools the agent may use, and which need a person's approval first. If a tool's definition changes on the server, the agent stops using it until you review it again.

Each mailbox has daily caps: by default 200 replies a day, 5 per conversation and 10 per recipient. Anyone can reply “STOP” to stop automated replies, and addresses that bounce or complain are suppressed.

Signed webhooks tell your app when mail arrives and what the agent did with it:

message.received   a new message, with its screening result
message.dropped    a message not stored because the mailbox is full
action.held        a reply or tool call waiting for a person
action.executed    a reply sent or a tool call run
action.failed      a reply or tool call that couldn't complete

Already working in Claude, ChatGPT, Cursor or your own code? Connect it to a mailbox over MCP. It reads, searches and replies with the same screening, approvals and limits. Set it up.

No account yet? An agent can sign itself up with one API call and get an inbox straight away.