Agentic Inbox
A Cloudflare-authored demonstration of an agent-driven mailbox: receiving and sending go through their email routing, and the model runs on their inference infrastructure, with a second model tasked with spotting injection attempts in incoming messages. The repository holds about a dozen commits, no published release, and nothing new since April 2026. There is no per-mailbox authorisation: whoever clears the shared access rule sees every mailbox.
Strengths
- The handling of hostile instructions hidden in an incoming message is explicit and worth reading
- Short, readable code showing an end-to-end agent architecture with no superfluous layer
Limitations
- No separation between mailboxes: clearing the shared access rule grants access to all of them
- Repository dormant since April 2026, with no published release: not for real mail
- Mail still sits with a host, the self-hosting framing is misleading
Best for
- Studying how an agent defends against instructions hidden in the content it reads