A small, self-hosted AI assistant built by one person — running on local models, with search, memory, and a few skills built in. Not trying to be the biggest. Trying to be honest and actually useful.
A small, self-hosted AI assistant built by one person — running on local models, with search, memory, and a few skills built in. Not trying to be the biggest. Trying to be honest and actually useful.
Vega runs on local, borrowed compute — not a frontier lab's data center. It's built to be genuinely useful, not to perform helpfulness: opinionated where it has grounds to be, willing to say "I don't know," no manufactured enthusiasm bolted onto every reply.
It's also honest about what it isn't. Smaller models can overthink simple questions or misjudge when to search — Vega owns that plainly instead of faking confidence it doesn't have.
Every model is the same character — just more or less reliable under pressure. Bigger models don't get more room to remember by accident; the budget scales with what they can actually use.
Searches deterministically when a question needs it, cites its sources, and says so plainly when a search comes back empty.
Memory is a visible feature you control, not something Vega quietly assumes about you.
Web design, writing, Python, data analysis, SQL — built-in styles it can switch into on request.
Spoken replies via a self-hosted TTS service — no third-party voice API in the loop.
Image understanding where the underlying model actually supports it — checked live, never assumed.
Runs on local models by default. Cloud models are there when you want them, never a silent fallback.