MailKite
Get your API key
All guides Agent email

Agent email with Cloudflare

Give an AI agent its own inbox and it can hold a real email conversation — field a support request, follow up on a thread, confirm a booking. This guide sets that up with Cloudflare Email Service and Email Routing, then shows the same thing the MailKite way, where the agent gets email as native tools over MCP.

What you'll need

  • A Cloudflare account with a domain on Cloudflare and Email Routing enabled.
  • A Worker to deploy — sending and inbound both run as Worker code.
  • An agent runtime that can call tools (function calling or MCP).

Part 1 — Give your agent an inbox with Cloudflare

On Cloudflare, email lives inside a Worker: you send with the Email Service send binding and receive by pointing Email Routing at an Email Worker that parses the raw MIME. Both halves are the agent's tools.

1. Send from the agent

Bind EMAIL in wrangler.toml, then wrap the send call as your agent's send_email tool. You build the MIME message yourself:

worker.js
// wrangler.toml
// send_email = [{ name = "EMAIL", destination_address = "agent@yourdomain.com" }]
import { EmailMessage } from "cloudflare:email";
import { createMimeMessage } from "mimetext";

export default {
async fetch(req, env) {
// Expose this as your agent's send_email tool
const msg = createMimeMessage();
msg.setSender({ name: "Agent", addr: "agent@yourdomain.com" });
msg.setRecipient("ada@example.com");
msg.setSubject("Re: your order #1042");
msg.addMessage({ contentType: "text/html", data: "<p>Shipped — tracking is attached.</p>" });

await env.EMAIL.send(
new EmailMessage("agent@yourdomain.com", "ada@example.com", msg.asRaw())
);
return new Response("sent");
},
};

2. Receive

Point Email Routing at a Worker so an inbound reply hits your agent's loop. The Worker gets the raw message — parse it with postal-mime and feed the result back into the model:

email-worker.js
// wrangler.toml — route inbound to an Email Worker:
// Email Routing → your-worker → this handler
import PostalMime from "postal-mime";

export default {
async email(message, env, ctx) {
// Parse the raw MIME stream into a clean object
const parsed = await PostalMime.parse(message.raw);

// Hand it to your agent's loop
agent.handle({
from: message.from,
subject: parsed.subject,
body: parsed.text,
});
},
};

That's an agent inbox on Cloudflare: deploy a Worker for send, route inbound to an Email Worker that parses MIME, and let your agent run the conversation.

Part 2 — The same, the MailKite way

MailKite is MCP-native, so there's no Worker to deploy and no MIME to parse — your agent speaks MCP from any runtime and gets email as first-class tools. Connect the server once and the same send-and-reply loop is a few tool calls.

1. Connect the MCP server

Add the hosted server. On the first connect it signs you in over OAuth in the browser — no key to copy — or pass your account key for headless runs:

connect
# Hosted MCP server — OAuth in the browser on first connect, no key to copy
claude mcp add --transport http mailkite https://mcp.mailkite.dev/mcp

# Headless / CI — pass your account key instead of the browser flow
claude mcp add --transport http mailkite https://mcp.mailkite.dev/mcp \
--header "Authorization: Bearer mk_live_..."

Any MCP client — Cursor, Claude Desktop, Cline, Zed — takes the same URL:

mcp config
{
"mcpServers": {
"mailkite": { "url": "https://mcp.mailkite.dev/mcp" }
}
}

2. Send from your agent

The agent calls mailkite_send over any verified domain — HTML, text, cc/bcc, attachments, and in-thread replies. Pick AI in the card for the agent's view, or any language to make the same send from your own code:

send
import { MailKite } from "mailkite";

const mk = new MailKite(process.env.MAILKITE_API_KEY);

const { id, status } = await mk.send({
from: "hello@myapp.ai",
to: "ada@example.com",
subject: "Your invoice #1042",
html: "<p>Thanks! Receipt attached.</p>",
});
Install Docs →
response
202 Accepted
{ "id": "msg_2Hk9…", "status": "queued" }

3. Receive — or hand the whole inbox to the agent

Every address on the domain is receivable. The agent can read inbound with its own tools, or you turn an address into an inbox agent with one route — the model reads and replies on its own:

create the agent route
import { MailKite } from "mailkite";

const mk = new MailKite(process.env.MAILKITE_API_KEY);

// One route turns an address into an AI inbox agent — the prompt
// is the program, and unresolved mail escalates to a human.
await mk.createRoute({
match: "support@myapp.ai",
action: "agent",
agentPrompt:
"Answer billing and account questions for myapp.ai. Reply in-thread. " +
"Escalate anything you can't resolve to humans@myapp.ai.",
agentForwardTo: ["humans@myapp.ai"],
});
Install Docs →
read inbound
// Or read inbound yourself — the agent has tools for it
mailkite_list_messages({ limit: 20 }) // newest first
mailkite_get_message({ id: "msg_2Hk9…" }) // full body, headers, attachment links

Same key manages the domain too — mailkite_create_domain, mailkite_verify_domain — so the agent can provision its own inbox end to end. See Connect your agent for the full tool list.

A little about MailKite

MailKite is programmable email for developers and AI agents. Connect the MCP server once and your agent can send over a verified domain, read its inbound as clean JSON, spin up new inboxes, manage domains, and become an inbox agent that answers mail on its own — all as native tools, no REST wrapper to maintain. It works the same over the SDKs, the CLI, or raw HTTP, and agent inboxes are on the free tier.

Next: Connect your agent for every MCP tool, or Set up your agent's email to go from zero to a working inbox.