Herald of the Gods

Since I started fossforus.watch I knew I would not have enough time to handle everything myself directly. I hoped (and still hope) to pick up some other volunteers along the way but that hasn’t happened yet. The need for more hands, however, is real and already felt.

So, inspired by the good people from Jupiter Broadcasting, I decided to try and automate the scouting for and publication of media articles.

A messenger, a “brain”, and a chat

The messenger is Hermes Agent, a free and open source agent framework by Nous Research. Hermes doesn’t “think”; it follows a “skill”, a plain markdown playbook I wrote that describes the whole job: fetch a curated list of RSS feeds, score each new item against the site’s selection criteria, and send me a shortlist.

The “brain”¹ doing the scoring is Gemma, an open-weight language model running on Ollama (self-hosted, like everything else here). No proprietary AI services, no data shipped to Big Tech clouds.

The chat is Signal, via the excellent signal-cli. Once a week the shortlist lands in my Note-to-Self, looking something like this:

📋 Weekly FOSSFORUS Digest — 5 candidates found

📰 [1] MEDIA · 9/10
*University X migrates 30,000 accounts to Nextcloud*
Some Outlet · 2026-07-13
🔗 https://...

Which numbers do you want to publish?

I reply with numbers or natural language. Hermes drafts each entry in the site’s format, shows it to me, and on my go-ahead pushes it to the Hugo site on Codeberg and toots it from @[email protected]. A human (myself) reads and approves every single thing before it goes anywhere.

From “0 of 0 articles found” to published

Things work surprisingly well — now. The first runs were humbling. The model would choke on too much input and confidently report there was nothing to report, half my feed URLs turned out to be dead, and the scoring kept surfacing “AI in the classroom” pieces exclusively. Each fix was small (feed the model less at a time, repair the feeds, sharpen the criteria), and now the whole pipeline works: scouted, scored, approved, published, tooted.

I also made it so if I find an article myself, I can just send the link to Hermes over Signal and ask it to publish. It fetches the piece, drafts the entry and the toot, and waits for my yes.

What’s next?

Fine tuning, mostly. The scoring criteria will keep being adjusted as I see what the digest gets right and wrong each Monday, and the feed list is a garden that needs prunning because sources may die, drift off-topic, or get replaced by better ones. Adding one is a couple of lines:

media:
  - name: "The Document Foundation"
    url: "https://blog.documentfoundation.org/feed/"
    type: "media"

And of course, the automation scouts, but it doesn’t read or think. If you’d like to be a human in this loop, let’s join forces.