A personal AI system: one agent and one app on every screen
Personal AI system The system I build for myself: the agent Will takes tasks by voice or text, hands them to other models, and nothing leaves without my «yes».

Overview
Vega is my own project, built for myself. Instead of a dozen apps and chats there is one counterpart, the agent Will, and one app on the laptop, the phone and the glasses. I say or type a task, Will hands it to other models, watches the work and comes back with the result. Inside is Vegapunk, my system of agents; Vega is its face.
A head and hands
Will is the head: it talks with me, writes briefs and accepts results. The work is done by the hands, separate model sessions: Codex writes code, Claude writes texts and designs, local models handle anything personal. Each task leaves with a goal, the files and the command that checks the result.
How a request goes
- Me
I say what I need, in my own words.
voice, text, glasses
- Will
Understands, writes the brief, picks the hands.
Claude
- Hands
Do the work and check it by command.
Codex · Claude · local models
- ✅ card
Anything that goes out waits for my fingerprint.
fingerprint or face
- Result
Will reports what is done and what is not checked.
one sentence or a page
The ✅ gate
Agents write, check and assemble on their own, but they send nothing in my name. A message to a person, a site release, a payment or a signature stops at a card, and only my fingerprint or face lets it through.


Work
All tasks and questions live in the Work section: projects, their sessions and questions to me. A question arrives with options and a recommendation, the answer is one tap. That is how I answer them without opening a single session.

My data stays on my machines
Anything personal, correspondence, health, money, is processed only by local models on my machines at home: a home server that recognises speech and a DGX Spark with a local model. Cloud models get work material only. Any model can be replaced if a new one does no worse on my real tasks.
Where things run
- Home server
Will, the Vega app, speech recognition and the voice, the archive.
Tesla T4 · RTX 2060
- DGX Spark
The local model: everything personal goes only here.
Qwen · Gemma
- Cloud models
Work material: code, texts, design, research.
Claude · Codex
One look on every screen
I designed the interface myself: one design system for the browser, the phone, the glasses, the watch and the TV. Orange means one thing only: it waits for me. Music, films, books and the calendar live in the same app, without feeds or ads.

In numbers
- 279voice messages a weekAll transcribed on my own machine.
- 141✅ cards a week94 approved, 13 declined.
- 28agent sessions at onceOn two machines at home.
- 19model rolesEach role is a chain of models; a model is replaced only after a test on my real tasks.
- 456questions answered a weekEach with options and a recommendation.
Credits
- Idea, architecture and interface
- Jean Ivanov
- Code: AI agents under Jean’s direction
- Claude, Codex
Next projectAn app that turns school into a gameBaitApp

