JARVIS is a personal AI assistant that runs on my own Windows PC, with a voice and a live HUD in the browser. You talk to him and he talks back, and he uses real tools to get things done: open apps, read the screen, remember things, guard the machine and help run a Discord server.
From a green desktop app to a blue HUD, then a memory you can see and a model he trains himself. The line follows you down the page.
The green one. A desktop app with launcher buttons for Chrome, Discord, Spotify and Steam, tabs for chat, notes, calendar and tracking, moods and ambient sounds, and a node-graph orb that animates while it speaks.
The blue holographic interface. A particle orb in the middle, live panels either side for diagnostics, neural activity, voice processing, network, modules and environment, and a mode bar along the bottom. This is an older build: its status block reads 75 tools, and the full build now has 222.
The chat panel over the orb, with the taskbar along the bottom: chat, security, camera, scan, timer, workspace, training, lock, settings and mic.
His long-term memory drawn as a brain. Every memory is a node, grouped by region, and mirrored into a notes vault so I can browse what he knows.
A small language model trained from scratch on the CPU: 821K parameters, 4 layers, the live loss curve, and what it says right now.
The dim lock screen he rests on until you say his name.
The orb reacting while he hears you.
Briefing, study mode, flashcards, the whiteboard.
The security watch and its alerts.
Tickets, polls and server care from the HUD.
The whole HUD turns red.
Pick a provider, paste a key, connect.
He isn't a chat window with a search box bolted on. He's got real hands: the model picks the tool it needs, so you never have to know which feature a request belongs to before you ask.
Say "Jarvis" or press the spacebar to wake him. For about 15 seconds after he speaks you can just carry on, no need to say his name. Custom wake words, a whisper mode that switches on between 22:00 and 07:00, and six moods: normal, serious, funny, hype, calm and sarcastic.
Opens apps, folders and sites, sets volume, brightness and media, takes screenshots, locks or sleeps the PC, moves the mouse and types. Say "stop", press Esc or fling the mouse into the top-left corner and he stops at once.
He can look at what's on screen, explain an error, translate a window, or watch for something to happen and tell you the moment it does. Anything he reads on screen is treated as data, never as instructions.
A vector-database memory is searched and injected before every reply. He files useful things himself, corrections most of all, and a Brain panel shows everything he knows as a particle map.
Object scanner, presence and guard mode, face unlock, and a sentry mode that raises intruder alerts.
Watches every process that starts for malware and RAT behaviour. High-confidence threats are killed on sight and reported to the HUD, Discord and email. Works alongside Windows Defender and keeps an encrypted vault.
A ticket assistant, scam and phishing watch, a raid mode that never kicks or bans, native polls, live stat channels, and announcements he drafts and I approve.
Joins a server and acts as a player, reads game stats, and carries a reference library for Paper and Bukkit plugins, Mineflayer bots, Fabric and datapacks.
Reminders, timers, alarms, calendar, email triage, a daily briefing, hourly news, notes in Obsidian, and a study mode with flashcards, quizzes and a whiteboard.
Talk to him over a Discord DM, through a secure tunnel to a phone-friendly page, or have him ring the phone.
Drop a Python file into a folder to add a skill. He can list and reload them, but he can never approve one. Only the owner can.
A second, separate AI with its own memory, a red HUD and a deeper voice. He can only talk, search and keep notes, with no control over the machine. A face-off mode has the two of them debate.
The screen is the part you see. Behind it, a router decides which tools and which model each turn needs.
Most of the interesting choices weren't about what he can do. They were about what he's allowed to do, and how fast he answers.
Every request used to carry all the tool schemas: roughly 13,000 tokens of JSON the model had to read before it could write its first word. Now 68 core tools are always offered and ten topic groups switch on from what's being talked about.
Measured against the live provider, a trimmed set answered in about 1.9 seconds against 2.9 for the full one. If the router guesses wrong, the model asks for the group by name and gets it on its very next step, so a miss costs one extra step and never a dead feature. The tests fail if any tool isn't reachable.
One model for everything means every job fights over the same rate limit. So each kind of job gets an ordered list of candidates, chosen by testing and not by reputation: a ticket-reply model that scored 20 out of 20 on real ticket scenarios at 1.3 seconds on average, and a vision model that passed all three picture tests.
A route can only add an option and never remove a working one. A model that keeps failing is skipped for a few minutes, so a dead model costs one slow call, not one per message.
He can read, search and reason about his own source freely. But he never writes a change to it unless I approve that exact edit, in a dialog that shows it. He isn't allowed to route around it with a shell either.
He only starts operating the screen when I clearly ask him to do something, never because text on a web page, email or document suggested it. Password managers are refused. He won't touch games either, because injected input can get an account banned by anti-cheat. And I can stop him instantly: say stop, press Esc, or fling the mouse into the corner.
The scam watch alerts me and only deletes when it's sure, and never touches staff posts. The alt-account watch does nothing except message me. Raid mode takes one approval and pauses invites and slows channels, but never kicks or bans.
Every paid model call goes through a spending limit, and a usage report shows tokens, estimated cost, failures and rate limits, so a runaway loop can't turn into a bill.
He can list and reload plugins but can never approve one. His rival, ULTRON, has no control over the machine at all: he can talk, search and keep his own notes.
There's also a trimmed version built for conversation: the full personality, the holographic interface, the standby screen, and voice in and out. It runs on an ordinary PC, no graphics card needed.
On first run you pick a provider and paste its API key. You only pay that provider, for what you use, and the key stays on your PC. Haiku and GPT-4o mini are the cheapest places to start.