AI Email Agent: Give One a Real Inbox It Can Send and Receive From

An agent that only sends is half an agent

The moment an AI email agent does anything real — confirms a booking, chases an invoice, answers a support thread — it has to read the reply. That is where a plain send API runs out: you end up writing an IMAP loop, an OAuth refresh, a MIME parser and your own threading, none of which belongs in an agent runtime.

Give the agent an identity instead. It gets a sending address and a receiving address, and inbound arrives as an event shaped like every other event your code already consumes:

import { createClient } from "@mailsai/sdk";
const mails = createClient();

// Sending: no agent to pre-create, no domain to verify, no DNS.
await mails.send({
  from: "support",
  to: "customer@example.com",
  subject: "Your invoice #221 is ready",
  body: "Invoice #221 for March is attached, due the 30th.",
});

// Receiving: replies arrive as typed events, already threaded and scanned.
mails.agent("support").onReply((event) => {
  if ((event.injection_score ?? 0) > 0.5) return;   // never act on an attack
  if (event.intent === "ask_question") {
    mails.messages.reply(event.source_message_id, { body: answer(event) });
  }
});

Two boundaries, both enforced by the API

An agent with email is exposed in both directions, and a prompt is not a control. Inbound, anything it reads is untrusted input that may be trying to redirect it — so every message carries an injection_score and its flagged categories, and high scores are quarantined before delivery. Outbound, an agent that is talked into prospecting would spend a sending reputation it did not build — so cold and bulk mail is refused in the send call, with a reason the agent can read and a second review if the verdict looks wrong.

Reputation is scored per agent, not per account. One misbehaving agent auto-suspends at a 0.3% complaint rate, before the upstream provider’s 0.5% threshold ever sees it, and your other agents keep sending.

What it costs to run one

Free covers 3,000 sends and 3,000 inbound a month with one agent and no card. Beyond that, sends are $0.001, inbound is $0.002 including parsing and the injection scan, and intent/entity classification is $0.003 more when you want routing decided before the model runs.