# Celebrating small wins
Sunday, 4. October 2026


About one month ago I released [Postern](https://code.raphting.dev/postern.git), an open-source Mail Delivery Agent with
a policy engine at its core.

I took two weeks off in September, and I am back at my day job at [p25.dev](https://p25.dev).

p25.dev's MTA is [KumoMTA](https://kumomta.com). I am happy with this choice, and I contributed [a](https://github.com/KumoCorp/kumomta/pull/490)
[few](https://github.com/KumoCorp/kumomta/pull/523) [PRs](https://github.com/KumoCorp/kumomta/pull/540) since.
My contribution was
[highlighted in the latest Changelog](https://github.com/KumoCorp/kumomta/blob/main/docs/changelog/2026.09.29-b90d8bc1.md#:~:text=raphting).
Thanks Wez for the kind collaboration!

p25.dev had some hiccups with Outlook (as every ESP has at least once), but I could sort it out fairly quickly with
their support agents. There seem to be real humans working, answering quickly, and that was a good experience.
(Don't use Outlook though ;) )

Postern's policy script can make HTTP requests now. Via HTTP, I can call my LLM provider Scaleway to generate a response
based on a prompt and email content. I want to classify a dedicated set of emails.
LLM is not the only but a convenient way to do that.

> Because I write about email classification: This is not about SPAM. The idea to classify emails for HAM and SPAM with
> LLM is old. The problem always was execution speed and cost. This might change with
> [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) or open
> "system one models" once there will be open-source parameter sets.

For Postern, I also added the option to call binaries from the policy script and pipe emails
(or attachments of those emails) into them.
I use this for an upload script for specific email attachments that I want to store separately.

I will soon push those changes and formally release the first tagged alpha version of Postern. From a functionality
perspective, it is sufficient, but I need to write more documentation so that users can get started easily.
A good amount of experimentation went into Postern's design to get the basic workflows smooth. Even if there's nobody
else interested in using it (which I hope is not the case), I am so hyped about it that I don't want to work with any
other MDA anymore.

Just wanted to celebrate a few small wins here.


By Raphael Sprenger