AgentMail Review: Email Inboxes for AI Agents (Setup, Pricing, Use Cases)

AgentMail is not an AI that files your Gmail. It gives AI agents their own inboxes: real addresses they can create, send from, receive on, and automate against through an API. That sounds small until you notice what agents cannot do without it — verify accounts, talk to vendors who only accept email, or run a loop that starts when a message arrives.

The company raised a $6 million seed round led by General Catalyst, with Y Combinator, Phosphor Capital, and angels including Dharmesh Shah, Paul Copplestone, and Karim Atiyeh. The product is live. In the video below, we connect it to Codex, walk through seven use cases, and build a daily check that emails you when a model is available again.

What AgentMail actually is

AgentMail is an API-first email provider for agents. You create inboxes programmatically, get addresses such as you@agentmail.to (custom domains are supported on paid plans), and send or receive with threading, labels, search, webhooks, and WebSockets.

That is the reverse of "AI for your inbox." Your agent becomes a mail user. Codex, Claude, OpenClaw, or anything that can call an API or MCP server can own the mailbox.

First-party writeup of the round and the thesis: We've raised $6M to build the first email inboxes for AI agents.

Pricing (September 2026)

From AgentMail pricing, current published plans are:

• Free — $0/month: 3 inboxes, 3,000 emails/month, 100 emails/day, 3 GB storage, 2 webhook endpoints. No credit card. Custom domains are not included.
• Developer — $20/month: 10 inboxes, 10,000 emails/month, no daily send cap on that plan's marketing copy, 10 GB storage, 10 custom domains, email support.
• Startup — $200/month: 150 inboxes, 150,000 emails/month, 150 custom domains, SOC 2 report, Slack support. Early-stage startups can apply for a free month on this tier.
• Enterprise — custom: unlimited inboxes available, EU or BYO cloud, dedicated IPs, SAML/OIDC SSO, white-label.

MCP server access is listed on every plan, including Free. Confirm limits in the console before you build a high-volume loop; the comparison table on the pricing page lists a 1,000 emails/day cap on Developer even though the plan blurb says no daily limit.

Setup inside Codex

The useful demo in the video is how little setup you do by hand:

• Create an API key in the AgentMail console.
• Ask Codex to install AgentMail from the docs. It reads the API docs, wires the MCP connection, and tells you what to paste in the terminal.
• Authorize once. After that, the agent can create inboxes, send, and list mail without you writing the integration.

If you are new to this pattern, start with MCP servers explained. AgentMail is a good first MCP because the actions are obvious: send, read, reply.

Seven use cases that are actually email-shaped

Skip newsletters. Dedicated broadcast tools still win there. Use AgentMail when the agent needs a mailbox as an interface:

• Automated support: inbound questions land in the agent inbox, the agent drafts or sends a reply, you only see escalations.
• Lead intake: a unique inbox per campaign or per agent, so you can retire it without touching your personal mail.
• Account verification and signups: the agent receives the confirmation email and continues the flow.
• Vendor coordination: anything that still insists on email instead of an API.
• Daily briefings: scheduled checks that email you a summary, then accept your reply as the next instruction.
• On-demand inboxes: spin up an address for a task, delete it when the task ends. That is the real unlock, not owning one forever inbox.
• Human-in-the-loop approvals: the agent emails you a proposed action and waits for a yes.

A concrete automation: email me when the model is back

In the video we ask Codex to watch whether Anthropic's Fable model is available and email the answer every morning. First test is even smaller: "email me a poem." Codex finds its AgentMail inbox, asks permission, and the message hits a personal inbox in seconds.

Then we reply to the automation email and tell Codex to check the AgentMail inbox — not Gmail — and respond. The only snag was the agent guessing Gmail the first time. Naming the inbox in the prompt fixed it. That is the product: email as a bidirectional UI between you and an agent, with webhooks or a scheduled check to make it autonomous.

If you are chaining this into a longer agent loop, see what are AI loops.

When not to use AgentMail

• You only need to draft replies inside your own Gmail. That is an inbox assistant, not an agent identity.
• You are sending marketing newsletters. Use a broadcast vendor.
• You cannot yet verify the domain or the From address your recipients will trust. Start on agentmail.to, then add a custom domain on Developer or higher before you email customers.

Bottom line

Agents that only chat in a browser are trapped. Agents that can own an inbox can talk to the rest of the internet the way companies already do. AgentMail is early infrastructure for that: cheap to try, MCP-friendly, and most useful when you treat inboxes as disposable resources rather than one more personal account.