Every morning, I spent 20 minutes checking AI news, live model pricing, and new releases across Hacker News, Techmeme, and Reddit.
It was repetitive, tedious, and hard to track.
So I decided to automate the entire thing. But looking at the monthly bills for hosted automation tools like Zapier or Make, I realized those costs scale fast. I wanted complete ownership of my data, schedules, and integrations, and I didn't want to pay monthly SaaS fees.
I built a completely free, self-hosted AI automation stack.
The Stack
The stack runs 24/7 on an always-free Oracle Cloud VPS. It coordinates:
- A Python research engine that scrapes news sources.
- OpenRouter free model routing to generate structured summaries.
- A Notion database tracking 18 properties.
- Instant Slack alerts containing the daily briefing highlights.
In this guide, I'll walk you through how you can own your automation stack too...
Repo Links:
Top comments (4)
Really like this — the "own it" framing is exactly why I got into self-hosting too (self-taught, I run mine for a small hospitality business, more bar-counter than compiler).
One honest flag on the ownership angle: OpenRouter free models are still an external dependency — free tiers change, rate limits move, and your prompts leave your box. If full ownership is the goal, the LLM step is actually the easiest part to pull fully local. I swapped mine for Ollama on my own machine, so nothing leaves the box and the cost stays a hard €0 even when I hammer it.
The trap that actually cost me time wasn't the hosting though — it was trusting the model's output. For anything structured (my case was reading invoices) I now force the model to return a confidence field and flag low-confidence items for review instead of silently writing a plausible-but-wrong value. A briefing that quietly invents a release or a price is worse than no briefing.
Genuinely curious: does your daily summary ever hallucinate a model/price that wasn't in the scraped source, and how are you catching it — or does it not really matter for a briefing?
Fair flag on OpenRouter — free tiers move and you're right that the LLM step is the easiest thing to pull fully local. I've kept it external for convenience, but an Ollama swap is a clean drop-in if you want nothing leaving the box.
On hallucination: you're right that it's a real risk — I don't pretend the model doesn't hallucinate. My defence is two-fold: (1) the prompt is hand-tuned as far as I can push it so output stays constant, and (2) every bullet in the briefing carries its source link straight from the scrape, so I can trace each claim back to the original article. That way a wrong claim is caught by the link, not buried.
That said, your confidence-field idea is genuinely the next step — forcing a low-confidence flag for review instead of silently writing a plausible-but-wrong value. Definitely stealing that one. Thanks for the honest engineering feedback!
TextStow could be useful for this workflow — clipboard history + reusable favorites + prompt templates + cleanup for JSON/PDF/URLs. Local-first, free: textstow.com
Oh nice, thanks for sharing this! I'm running on Windows so I can't test it myself since it's Mac-only, but a local-first clipboard manager fits the self-hosted vibe perfectly. Definitely a cool tool for the Mac devs reading this to check out.