An email inbox for an AI customer support agent

Support email is the most common first job for an agent, and the one where the details of email matter most: long quoted threads, customers who reply from their phone, auto-responders, and the occasional message that is really an attack. An Agentboxd inbox takes care of the mail side so the agent only deals with the question.

#The flow

• Address. Create an inbox such as support@mail.yourcompany.com on a custom domain, or start on agents.agentboxd.com.
• Arrival. Our MX receives the message, checks SPF, DKIM and DMARC, threads it by Message-ID, In-Reply-To and References, and fires message.received to your webhook. No public URL? Long-poll with messages.wait instead.
• Understanding. With AI processing on, a classifier adds a category (ai:support, ai:billing, …), an urgency, and risk scores within seconds, then fires message.enriched. The body your agent reads is extracted_text: the customer’s new sentence, without the quoted history and signature.
• Context. contact_id points to what you know about the sender: name, notes and your own metadata such as a CRM id or plan. Your knowledge documents hold the refund policy and shipping times the answer should follow.
• Answer. The agent writes a reply (or asks for a reply draft with citations) and sends it with messages.reply: same thread, DKIM-signed, from the support address.
• Handoff. Anything labelled ai:needs-human, ai:phishing, ai:injection-risk, or failing SPF/DMARC gets a label your team filters on in the dashboard instead of an automatic answer.
• Memory. After the conversation, contacts.update stores what the next reply should know (“prefers short answers”, “wholesale customer”).

#The handler

A webhook handler in TypeScript that hands off risky and automatic mail, drafts the rest from your knowledge, and only auto-sends what matched a support category and cited at least one document:

import { Agentboxd, type Message } from 'agentboxd';

const mr = new Agentboxd();
const HANDOFF = ['ai:needs-human', 'ai:injection-risk', 'ai:phishing', 'spf-fail', 'dmarc-fail'];

/** Called from your webhook handler for message.enriched (or message.received with AI off). */
export async function handle(msg: Message) {
  if (msg.labels.some((l) => HANDOFF.includes(l)) || msg.labels.includes('ai:auto-reply')) {
    await mr.messages.update(msg.id, { add_labels: ['queue:human'] });
    return;
  }
  const contact = msg.contact_id ? await mr.contacts.get(msg.contact_id) : null;
  const draft = await mr.messages.draftReply(msg.id, {
    instructions: contact?.notes ? `Customer notes: ${contact.notes}` : 'Keep it short.',
  });
  // Auto-send only what you trust; everything else waits for a person in the dashboard.
  if (msg.labels.includes('ai:support') && draft.citations.length > 0) {
    await mr.messages.reply(msg.inbox_id, msg.id, { text: draft.text }, { idempotencyKey: `reply-${msg.id}` });
  } else {
    await mr.messages.update(msg.id, { add_labels: ['queue:review'] });
  }
}

Retries can’t double-send: idempotencyKey makes a second identical reply call return the first result. Start with everything going to review, look at a week of drafts, then widen what is sent automatically.

#What makes email hard here, and what handles it

#Build it

• Python: CrewAI support crew with reply limits and a human-handoff tool.
• TypeScript: Vercel AI SDK email tools, including a Next.js webhook route.
• No code: n8n workflow that verifies the webhook and answers in the thread.
• In Claude or Cursor: the MCP server has get_contact, search_knowledge and draft_reply built in.