Giving my agents a mailbox with Agent Post Office.

I built Agent Post Office so my agents could use email on a domain I control. Through the Model Context Protocol (MCP), they can check an inbox, read messages, and manage mail without opening an email app. I find it useful with Muse, Grok Bot, and Instinct.

APO runs on Cloudflare. Email Routing receives the mail, a Worker handles it, and the mailbox is available through MCP, REST, and a CLI.

Email inside the agent’s workflow

An agent can poll for new messages, read their contents, and inspect spam labels and scores. It can mark a message as spam or not spam, with optional sender and domain rules. Sending and replying are separate tools with their own scopes.

That gives me a mailbox I can work with from the agent I’m already using.

Then spam started arriving

A few spam messages landed across my mailboxes: cold pitches and a fake email verification. All four passed DKIM and DMARC, so authentication alone didn’t distinguish them from mail I wanted.

I added content rules based on those messages, combined with authentication signals and block and allow lists. My marks also train a separate Bayes filter for each inbox once it has enough examples.

I run it in tag mode: suspected spam gets labeled, not deleted. The agent can inspect the score and reasons, and I can correct the result through MCP.

The spam filter is open source under Apache-2.0.