Show HN: Skales – Local AI agent desktop app (.exe/.dmg, 300MB idle RAM)
I'm Mario from Vienna. I’m the designated "IT guy" for my family (the guy they call to plug in a Scart cable or to setup Chrome).
A few weeks ago, I spent over two hours wrestling with Docker (and some for my at this time unknown Terminal commands) just to get a CLI-based AI agent running. It ate all my RAM, and I had a realization: my wife, my non-technical clients, or my 6-year-old son could never set this up. If I deployed something like this, I'd spend my life doing setup-support (and this isn't just about finding the HDMI 1 to HDMI 2 input on the remote control).
I’ve been "vibecoding" with LLMs for a while the last 2 years, mostly building common landingpages or simple retro games to show my son that we live in a wild time where you can build a lot of 'entertaining' stuff just by prompting. I decided to pivot an old, failed Laravel SaaS of mine (Skales) into a native desktop app to solve this setup hell.
The only goal I wanted to reach: an autonomous AI agent that installs like a normal app (.exe / .dmg). No PhD required. Download, Install, Done.
It actually worked. My 60+ year old mom got it running instantly (so.. is Skales now a Grandma-Approved AI Agent?), and my 6yr used the built-in Coding Skill to make his own retro game (I wouldn't say its really a 'game' when it's basically a one level copy of Super Mario - but he loves it).
What it does:
ReAct autopilot, bi-temporal memory, browser automation (Playwright), and native integrations (Gmail, Telegram, Calendar).
BYOK: Works with OpenRouter, OpenAI, Anthropic, or local Ollama.
Tech stack: Electron + Next.js + Node.js (I managed to keep idle RAM around ~300MB).
Everything is stored locally in ~/.skales-data.
I know that being a "design guy who vibecoded an Electron app" (or Marketing-Dude) is basically fresh meat for the HN crowd. Feel free to roast the architecture! But my main goal was just making these incredible AI tools accessible for everyday tasks (formatting resumes, sending PDFs, building simple games) without touching a terminal or moving from one tool to another tool.
License note: The project is BSL-1.1 (Source-available, free for personal use). I chose this simply because I don't want a giant company to just take the repo, build their brand around Skales and commercially resell it, but I want the code out there for the community to use and learn from.
(excuse my English - i'm not a native speaker)
Would love your honest feedback on the UX (could be maybe better?)!
GitHub: https://github.com/skalesapp/skales Website: https://skales.app
For Execute Mode (multi-step autonomous tasks), Skales queues steps sequentially so there's no parallel bottleneck – it plans, you approve, then it runs through each step. There's also a desktop buddy (think as Microsofts Clippy, but actually useful) that sits in your system tray (if activated) as soon you minimize or close the main-windows, you can ask it quick questions without even opening the main interface. It runs within the same Electron process, so zero additional RAM overhead. Idle RAM sits around ~300MB (I had 400MB at least) which keeps things snappy.
The main speed factor is honestly the LLM provider, not Skales itself. With local Ollama models it's purely your hardware.
Happy to answer more specific questions, thank you for asking jlongo78!
https://www.youtube.com/watch?v=8fXGsQGyxCU
One thing worth considering as you build out the agent UX: the quality of the default prompts/instructions you ship with Skales will matter a lot for first impressions. A non-technical user can't debug a bad system prompt — they'll just think the AI is dumb. Structuring those instructions carefully (role, constraints, examples) makes a huge difference.
Built flompt (https://flompt.dev / https://github.com/Nyrok/flompt) for exactly that — visual prompt structuring that helps get the instructions right before they become a UX liability.
Thanks for the thoughtful feedback!
The BSL-1.1 license choice is interesting too — shows you're thinking about sustainability and intent, not just shipping fast. That kind of explicit decision-making is actually what separates vibe coding projects that survive from ones that collapse: when the human behind the project is thinking about architecture, maintainability, and accountability.
Somewhat related: the Agile Vibe Coding Manifesto (https://agilevibecoding.org) is trying to formalize exactly these principles — that customer value and human accountability still drive everything even when AI is writing most of the code. Your project is a good example of vibe coding done with intention.
Good luck with Skales — the accessibility angle (no Docker, no CLI) is genuinely underserved.
Also launched on Product Hunt today: https://www.producthunt.com/products/skales - would appreciate any feedback.