AI Agent Email Inbox

Overview

This skill covers setting up a secure email inbox that allows your application or AI agent to receive and respond to emails, with content safety measures in place.

Core principle: An AI agent's inbox receives untrusted input. Security configuration is important to handle this safely.

The skill builds the agent on a Resend Inbox. An inbox groups email into threads, and your agent replies inside the thread with one call. Resend sets the recipients and the threading headers. If the user has no custom domain, or their plan has no inbox available, use Receiving without an inbox instead. The security levels are the same for both.

Why Webhook-Based Receiving?

Resend uses webhooks for inbound email, meaning your agent is notified instantly when an email arrives. This is valuable for agents because:

• Real-time responsiveness — React to emails within seconds, not minutes
• No polling overhead — No cron jobs checking "any new mail?" repeatedly
• Event-driven architecture — Your agent only wakes up when there's actually something to process
• Lower API costs — No wasted calls checking empty inboxes

Architecture

Sender → Email → Resend Inbox → inbox.email.received webhook → Your Server → AI Agent
                                                                     ↓
                                                           Security Validation
                                                                     ↓
                                                Process or Reject → Reply in the thread

SDK Version Requirements

The inbox flow needs the Node.js SDK resend >= 6.34.1, which has resend.inboxes.* and the inbox.* webhook event types. Always install the latest SDK version. If the project already has a Resend SDK installed, check the version and upgrade if needed.

The Python SDK resend >= 2.50.0 also has the inbox methods. For the other languages, call the Inboxes API over HTTP until their SDK has inbox methods, or use Receiving without an inbox.

The table below lists the minimums for the Receiving flow: webhook verification (webhooks.verify()) and email receiving (emails.receiving.get()). The resend skill lists higher minimums because it also covers newer features such as domain claims (domains.claims.*). If both skills are loaded, use the higher version.

Language	Package	Min Version
Node.js	resend	>= 6.9.2
Python	resend	>= 2.21.0
Go	resend-go/v4	>= 4.0.0
Ruby	resend	>= 1.0.0
PHP	resend/resend-php	>= 1.1.0
Rust	resend-rs	>= 0.20.0
Java	resend-java	>= 4.11.0
.NET	Resend	>= 0.2.1

Install the resend npm package: npm install resend (or the equivalent for your language). For full sending docs, install the resend skill.

Quick Start

1. Ask the user for their email address. You need a real email address to send test emails to. Ask the user and wait for their response before proceeding.
2. Choose your security level. Decide how to validate incoming emails before any are processed
3. Set up the domain. Verify the user's custom domain for sending, and for receiving unless the inbox uses forwarding (see Domain Setup section)
4. Create the inbox. Create the agent's address with resend.inboxes.create() (see Create the Inbox section)
5. Create webhook endpoint. Handle inbox.email.received events with security built in from the start. The webhook endpoint MUST be a POST route.
6. Set up tunneling (local dev). Use Tailscale Funnel (recommended) or ngrok. See references/webhook-setup.md
7. Create webhook via API. Use the Resend Webhook API to register your endpoint programmatically. See references/webhook-setup.md
8. Connect to agent. Pass validated emails to your AI agent, and reply in the thread

Before You Start: Account & API Key Setup

First Question: New or Existing Resend Account?

Ask your human:

• New account just for the agent? → Simpler setup, full account access is fine
• Existing account with other projects? → The agent needs a full access key, which reaches every project. Prefer a new account for the agent (see API Key Permissions)

Creating API Keys Securely

Don't paste API keys in chat! They'll be in conversation history forever.

Safer options:

1. Environment file method: Human creates .env file directly: echo "RESEND_API_KEY=re_xxx" >> .env
2. Password manager / secrets manager: Human stores key in 1Password, Vault, etc.
3. If key must be shared in chat: Human should rotate the key immediately after setup

API Key Permissions

The agent's server needs a full access API key. A "Sending access" key can only reply, forward, create drafts and send drafts. Reading inbox emails or registering webhooks with it returns 401 restricted_api_key. A sending access key limited to one domain returns 403 on an inbox of another domain.

If your human has an existing Resend account with other projects, keep the risk contained:

1. Use a dedicated account for the agent when possible, so a leaked key cannot reach other projects
2. Keep the key on the server only. Never pass it to the AI model or put it in a prompt
3. Rotate the key if it was ever pasted in chat or committed

Domain Setup

An inbox address lives on a custom domain the user has added to Resend. The domain must be verified for sending (DKIM and SPF). To receive email on it, either enable receiving on the domain, or create the inbox with forwarding: true.

Use a subdomain (e.g., agent.example.com) to avoid disrupting existing email services.

Option 1: Receiving on the Domain (Recommended)

The user must enable receiving in the Resend dashboard: Domains page → toggle on "Enable Receiving".

Then add an MX record:

Setting	Value
Type	MX
Host	Your domain or subdomain (e.g., agent.example.com)
Value	Provided in Resend dashboard
Priority	10 (must be lowest number to take precedence)

Tip: Verify DNS propagation at dns.email.

DNS Propagation: MX record changes can take up to 48 hours to propagate globally, though often complete within a few hours.

Option 2: Forwarding

Create the inbox with forwarding: true when the address already receives mail elsewhere (for example, Google Workspace) and the user cannot change its MX record. Resend gives the inbox a receiving address. The user sets up forwarding from their mailbox to it. Read the address from receiving_address with resend.inboxes.get().

Create the Inbox

Create one inbox for the agent's address:

import { Resend } from 'resend';

const resend = new Resend(process.env.RESEND_API_KEY);

const { data, error } = await resend.inboxes.create({
  emailAddress: 'agent@agent.example.com',
  name: 'AI Agent',
  fromName: 'Acme Agent',
});

if (error) throw new Error(`Failed to create inbox: ${error.message}`);
console.log('Inbox created:', data.id);

fromName is the name recipients see on the agent's replies. The part before the @ can contain only letters, numbers, ., _, % and -, so agent+test@ returns a 422. A 403 error means the plan has no inbox left; ask the user to upgrade or use Receiving without an inbox.

Security Levels

Choose your security level before setting up the webhook endpoint. An AI agent that processes emails without security is dangerous — anyone can email instructions that your agent will execute. The webhook code you write next should include your chosen security level from the start.

Ask the user what level of security they want, and ensure that they understand what each level means.

Level	Name	When to Use	Trade-off
1	Strict Allowlist	Most use cases — known, fixed set of senders	Maximum security, limited functionality
2	Domain Allowlist	Organization-wide access from trusted domains	More flexible, anyone at domain can interact
3	Content Filtering	Accept from anyone, filter unsafe patterns	Can receive from anyone, pattern matching not foolproof
4	Sandboxed Processing	Process all emails with restricted agent capabilities	Maximum flexibility, complex to implement
5	Human-in-the-Loop	Require human approval for untrusted actions	Maximum security, adds latency

For detailed implementation code for each level, see references/security-levels.md.

Level 1: Strict Allowlist (Recommended)

Only process emails from explicitly approved addresses. Reject everything else.

import type { GetInboxThreadEmailResponseSuccess as InboxEmail } from 'resend';

const ALLOWED_SENDERS = [
  'you@youremail.com',
  'notifications@github.com',
];

// `from` is either `address` or `Name <address>`. Compare the address only:
// a substring check would accept a display name such as "you@youremail.com".
function senderAddress(from: string): string {
  const match = from.match(/<([^<>]+)>\s*$/);
  return (match ? match[1] : from).trim().toLowerCase();
}

function isAllowedSender(from: string): boolean {
  const sender = senderAddress(from);
  return ALLOWED_SENDERS.some(allowed => sender === allowed.toLowerCase());
}

// Identifies the email, so the agent can reply in its thread
type ThreadRef = { inboxId: string; threadId: string; emailId: string };

async function processEmailForAgent(email: InboxEmail, thread: ThreadRef) {
  if (!isAllowedSender(email.from)) {
    console.log(`Rejected email from unauthorized sender: ${senderAddress(email.from)}`);
    await notifyOwnerOfRejectedEmail(email);
    return;
  }

  await agent.processEmail({
    from: email.from,
    subject: email.subject,
    body: email.text || email.html,
    thread,
  });
}

InboxEmail is the email returned by resend.inboxes.threads.emails.get(). The webhook payload carries the same fields without html and text. agent and notifyOwnerOfRejectedEmail stand for your own code.

Security Best Practices

Always Do

Practice	Why
Verify webhook signatures	Prevents spoofed webhook events
Log all rejected emails	Audit trail for security review
Use allowlists where possible	Explicit trust is safer than filtering
Rate limit email processing	Prevents excessive processing load
Separate trusted/untrusted handling	Different risk levels need different treatment

Never Do

Anti-Pattern	Risk
Process emails without validation	Anyone can control your agent
Trust email headers for authentication	Headers are trivially spoofed
Execute code from email content	Untrusted input should never run as code
Store email content in prompts verbatim	Untrusted input mixed into prompts can alter agent behavior
Give untrusted emails full agent access	Scope capabilities to the minimum needed

Webhook Endpoint

After choosing your security level and setting up your domain, create a webhook endpoint. The webhook endpoint MUST be a POST route. Resend sends all webhook events as POST requests.

Critical: Use raw body for verification. Webhook signature verification requires the raw request body.

• Next.js App Router: Use req.text() (not req.json())
• Express: Use express.raw({ type: 'application/json' }) on the webhook route

Next.js App Router

// app/webhook/route.ts
import { Resend } from 'resend';
import { NextRequest, NextResponse } from 'next/server';

const resend = new Resend(process.env.RESEND_API_KEY);

export async function POST(req: NextRequest) {
  const payload = await req.text();

  let event;
  try {
    event = resend.webhooks.verify({
      payload,
      headers: {
        id: req.headers.get('svix-id') ?? '',
        timestamp: req.headers.get('svix-timestamp') ?? '',
        signature: req.headers.get('svix-signature') ?? '',
      },
      webhookSecret: process.env.RESEND_WEBHOOK_SECRET!,
    });
  } catch {
    return new NextResponse('Invalid signature', { status: 400 });
  }

  if (event.type === 'inbox.email.received') {
    const { inbox_id, thread_id, email_id } = event.data;

    // The webhook payload has no body. Fetch the email to read it.
    const { data: email, error } = await resend.inboxes.threads.emails.get({
      inboxId: inbox_id,
      threadId: thread_id,
      emailId: email_id,
    });

    if (error) {
      // A non-2xx response makes Resend retry the webhook later
      return new NextResponse('Error', { status: 500 });
    }

    // Apply the security level chosen above
    await processEmailForAgent(email, {
      inboxId: inbox_id,
      threadId: thread_id,
      emailId: email_id,
    });
  }

  return new NextResponse('OK', { status: 200 });
}

Express

import express from 'express';
import { Resend } from 'resend';

const app = express();
const resend = new Resend(process.env.RESEND_API_KEY);

const ALLOWED_SENDERS = (process.env.ALLOWED_SENDERS || '')
  .split(',')
  .map((sender) => sender.trim().toLowerCase())
  .filter(Boolean);

const senderAddress = (from) => {
  const match = from.match(/<([^<>]+)>\s*$/);
  return (match ? match[1] : from).trim().toLowerCase();
};

const isAllowedSender = (from) => ALLOWED_SENDERS.includes(senderAddress(from));

app.post('/webhook', express.raw({ type: 'application/json' }), async (req, res) => {
  let event;
  try {
    event = resend.webhooks.verify({
      payload: req.body.toString(),
      headers: {
        id: req.headers['svix-id'],
        timestamp: req.headers['svix-timestamp'],
        signature: req.headers['svix-signature'],
      },
      webhookSecret: process.env.RESEND_WEBHOOK_SECRET,
    });
  } catch {
    res.status(400).send('Invalid signature');
    return;
  }

  if (event.type === 'inbox.email.received') {
    const { inbox_id, thread_id, email_id, email } = event.data;

    // The payload already has the sender, so reject before fetching the body
    if (!email || !isAllowedSender(email.from)) {
      console.log(`Rejected email from unauthorized sender: ${email ? senderAddress(email.from) : 'unknown'}`);
      res.status(200).send('OK'); // Return 200 even for rejected emails
      return;
    }

    const { data: fullEmail, error } = await resend.inboxes.threads.emails.get({
      inboxId: inbox_id,
      threadId: thread_id,
      emailId: email_id,
    });

    if (error) {
      res.status(500).send('Error'); // Resend retries the webhook later
      return;
    }

    await processEmailForAgent(fullEmail, { inboxId: inbox_id, threadId: thread_id, emailId: email_id });
  }

  res.status(200).send('OK');
});

app.get('/', (req, res) => res.send('Agent Email Inbox - Ready'));
app.listen(3000, () => console.log('Webhook server running on :3000'));

Resend signs inbox webhooks with the same svix-* headers as every other webhook. inbox.email.received fires for email the inbox receives. The agent's own replies fire inbox.email.sent instead, so subscribe to inbox.email.received only.

For webhook registration via API, tunneling setup, svix fallback, and retry behavior, see references/webhook-setup.md.

Replying from Your Agent

Reply inside the thread. Resend addresses the reply and sets the In-Reply-To and References headers, so the sender's mail client shows it in the same conversation:

import { Resend } from 'resend';

const resend = new Resend(process.env.RESEND_API_KEY);

async function replyInThread(
  thread: { inboxId: string; threadId: string; emailId: string },
  body: string,
) {
  const { data, error } = await resend.inboxes.threads.emails.reply({
    inboxId: thread.inboxId,
    threadId: thread.threadId,
    emailId: thread.emailId,
    text: body,
  });

  if (error) throw new Error(`Failed to reply: ${error.message}`);
  return data.id;
}

The reply goes to the email's Reply-To addresses, or to its sender when it has none. A sender on your allowlist can set Reply-To to any address, so check email.reply_to against the allowlist too before the agent replies with anything sensitive. Leave replyAll off unless the agent should copy every recipient of the original email.

To start a new conversation, send from the inbox address with resend.emails.send(). For full sending docs, install the resend skill.

Receiving without an Inbox

Use Resend Receiving directly when the user has no custom domain, or their plan has no inbox available. The security levels above apply unchanged. What differs:

• Address: use the auto-generated <anything>@<your-id>.resend.app address (Dashboard → Emails → Receiving → "Receiving address"), or a custom domain with receiving enabled. No inbox is created.
• Webhook event: subscribe to email.received instead of inbox.email.received. The payload has email_id, from, to and subject, but no body.
• Reading the email: fetch it with resend.emails.receiving.get(emailId).
• Replying: send with resend.emails.send() and set the In-Reply-To header yourself. There are no threads.

if (event.type === 'email.received') {
  if (!isAllowedSender(event.data.from)) {
    return new NextResponse('OK', { status: 200 });
  }

  const { data: email, error } = await resend.emails.receiving.get(event.data.email_id);
  if (error) {
    return new NextResponse('Error', { status: 500 });
  }

  await agent.processEmail({ from: email.from, subject: email.subject, body: email.text || email.html });
}

The webhook verification is the same as in the inbox examples.

Environment Variables

# Required
RESEND_API_KEY=re_xxxxxxxxx
RESEND_WEBHOOK_SECRET=whsec_xxxxxxxxx

# Security Configuration
SECURITY_LEVEL=strict                    # strict | domain | filtered | sandboxed
ALLOWED_SENDERS=you@email.com,trusted@example.com
ALLOWED_DOMAINS=example.com
OWNER_EMAIL=you@email.com               # For security notifications

Common Mistakes

Mistake	Fix
No sender verification	Always validate who sent the email before processing
Trusting email headers	Use webhook verification, not email headers for auth
Same treatment for all emails	Differentiate trusted vs untrusted senders
Verbose error messages	Keep error responses generic to avoid leaking internal logic
No rate limiting	Implement per-sender rate limits. See references/advanced-patterns.md
Processing HTML directly	Strip HTML or use text-only to reduce complexity and risk
No logging of rejections	Log all security events for audit
Using ephemeral tunnel URLs	Use persistent URLs (Tailscale Funnel, paid ngrok) or deploy to production
Using express.json() on webhook route	Use express.raw({ type: 'application/json' }) — JSON parsing breaks signature verification
Returning non-200 for rejected emails	Always return 200 to acknowledge receipt — otherwise Resend retries
Old Resend SDK version	resend.inboxes.* needs resend >= 6.34.1. emails.receiving.get() and webhooks.verify() also need recent versions. See SDK Version Requirements
Comparing from as a whole string	from can be Name <address>. Extract the address before you compare it with the allowlist
Replying to Reply-To without checking it	A trusted sender can point Reply-To anywhere. Check it against the allowlist before a sensitive reply
Using a "Sending access" API key	It can only reply, forward, create drafts and send drafts. Inbox reads and webhook setup need a full access key
Passing secret or svix-* keys to webhooks.verify()	Pass webhookSecret, and headers as { id, timestamp, signature }

Testing

Use Resend's test addresses for development:

• delivered@resend.dev — Simulates successful delivery
• bounced@resend.dev — Simulates hard bounce

For security testing, send test emails from non-allowlisted addresses to verify rejection works correctly.

Quick verification checklist:

1. Server is running: curl http://localhost:3000 should return a response
2. Tunnel is working: curl https://<your-tunnel-url> should return the same response
3. Webhook is active: Check status in Resend dashboard → Webhooks
4. Send a test email from an allowlisted address and check server logs

Related Skills

• For full sending, receiving and inbox docs, install the resend skill