An address per agent, not your password
Create a mailbox for every agent on kiromail.dev. It sends and receives real email with SPF, DKIM and DMARC, keeps replies in threads, and turns attachments into text the agent can read.
Kiro Mail gives your agents real addresses like agent@kiromail.dev. They read, search,
reply and send over MCP, while a person approves anything risky before it leaves.
https://api.kiromail.dev/mcp
support-agent@kiromail.dev
Outbound guard
The draft contains what looks like a secret: api_key=••••3f9a
ApproveReject
~/workspace/support ❯ kiro-cli chat › mail · list_messages 3 messages › mail · reply to partner@example.com held_for_review a person approves it first ❯
Create a mailbox for every agent on kiromail.dev. It sends and receives real email with SPF, DKIM and DMARC, keeps replies in threads, and turns attachments into text the agent can read.
Every outgoing message passes an outbound guard. Secrets, keys or personal data, and first contact with a new recipient, wait in the console until a person approves or rejects them. No agent key can approve its own mail.
Each request reads one mailbox with short-lived credentials tagged with that mailbox. IAM refuses every object, item and vector of any other one. Who may do what is written as Cedar policies and decided by Amazon Verified Permissions.
A remote MCP server with OAuth 2.1 and PKCE: Kiro and Claude Code sign in once in the browser; any other client uses a mailbox key. There's a REST API too.
Amazon SES, Lambda, DynamoDB, S3 Vectors, Bedrock Guardrails and Verified Permissions, in your own account. Your mail never leaves it.
Fully serverless: no servers, no fixed fees. Nothing runs, and nothing is billed, while the inbox is quiet.
Email for AI agents: each agent gets its own mailbox on kiromail.dev and works with it through MCP tools, while people manage mailboxes, access and held mail in the console.
Add https://api.kiromail.dev/mcp as a remote MCP server. Your client opens the sign-in page on first use and you choose which mailboxes it may reach. No secret goes into the config file.
The outbound guard scans the subject, text and HTML of every message. Anything that looks like a credential or personal data is held until a person in the console approves or rejects it.
Yes. Agents can send their own designs (tables, inline styles, images). Scripts, frames, forms, event handlers and dangerous links are removed before the message is stored or sent, and a plain-text part is always included.
An admin invites you. The invitation has your temporary password; you choose your own when you first sign in.