I cancelled Superhuman Mail

I cancelled my Superhuman Mail subscription. Not because the product was useless, but because I no longer wanted my mail workflow locked inside a polished SaaS client when the same work can sit on a harness I control.

Mail is a good test of that idea. It is noisy, multi-account, full of sorting rules, snoozes, and habits that accumulate over years. A model alone does not solve that. You need tools, IMAP, routing, memory, and a loop that can keep doing the work without you babysitting every message. And for me the decisive part is privacy: everything runs on my infrastructure, so nothing about this mail path is exposed externally.

Privacy first, on my own iron

That is the part SaaS mail clients never really solve. The moment sorting, snoozing, and agent decisions leave your network, you are trading convenience for someone else’s view into your inbox. Here the stack stays at home. Hermes, Qwen, the buckets, the learning loop, the dashboard, the crons. All of that lives on my side of the wire.

Nous Hermes Agent is the harness for that. It is model-agnostic, it has a real tool and skills layer, and it can work with mail over ordinary IMAP and SMTP.

On my Homelab it runs with Alibaba’s Qwen 3.8 27B on my NVIDIA DGX Spark. The Spark is a local Grace Blackwell box built for agent workloads, and Qwen 3.8 27B is an open-weights model I can run there instead of shipping every mail decision to a remote API. NVIDIA even publishes a Hermes-on-Spark path, so this pairing is not an accident.

Own the harness so you are not dependent on a certain model. Same idea as with Pi or Open Code, just applied to the inbox, and with the privacy benefit of keeping the whole loop local.

Semantic sorting, not dumb rules

In practice it already handles 5 mail accounts and sorts into 8 custom buckets. The analysis is semantic. Sender, subject, and body content are taken into account, and the folders themselves have a semantic meaning, so the sorting is based upon what the mail actually is, not upon a brittle keyword list that dies the first time someone changes a subject line.

That is the difference between an agent in a harness and a filter dialog from 2009.

The learning loop and the dashboard

The self-learning loop watches when I move mails into other folders and learns from that. When it needs an explanation, it asks me through Hermes Agent instead of silently guessing forever.

Hermes with Qwen also created a dashboard where I can sort mails manually, and those manual moves feed the same learning loop. So corrections are not a separate training project. They are just me using the system, and the system getting better because of it.

Snooze as folders, two crons underneath

Snooze is implemented on IMAP with special folders that mean time, like 1hour, 1week, or nextFriday. Mail goes into those folders when it should disappear for a while, and comes back when that meaning is due.

Under the hood, two crons keep the whole thing alive. One fetches, analyzes, and sorts mail. The other handles the snooze operations. Boring, mechanical, reliable, and mine.

Next step: my own mailservers again

The agent stack is already local. The next step is closing the remaining gap and running my own mailserver(s) again. Right now Hermes still talks IMAP into accounts that live elsewhere. Put the servers back under my roof as well, and the whole path from store to sort to snooze stays private end to end.

Why this is the right move for me

Superhuman was a subscription around a client. Hermes plus a local model is a stack around the work itself. I still care about speed and a clean inbox, but I care more about privacy and whether the agent runs in an environment I control, with open weights I can actually host, and with IMAP as the interface instead of a proprietary mailbox product as the ceiling.

I am not claiming this is zero dependency yet, and I am not claiming every inbox should look like mine. I am saying the decisive question is ownership of the harness and the place where it runs. Once that is settled, cancelling the SaaS client stops feeling dramatic.

Life is good.