Izuki — the AI companion that lives on your Windows screen
Talk to it. It sees your screen, answers you out loud, and uses your mouse and keyboard to get things done.
Free · Open source · Hands-free voice · Works with free AI keys
⬇ Download for Windows ·
🌐 Website — try it in your browser ·
📱 Phone app ·
How to use it ·
Support the builder ☕
macOS and Linux versions are coming soon.
Follow Izuki — @izukiapp everywhere:
X ·
Instagram ·
TikTok ·
YouTube ·
LinkedIn ·
All links
v1.0.66 — verified hotfix and a movable Island
The Island now has a drag grip and remembers its safe position along the top of the screen. Its aligned quick-action grid adds Screenshot and PC Boost. MIDI song drafts no longer reject a valid drum track when two hits share a pitch, model-requested waits are honored before screen verification, and update cards remain visible with accurate available or ready to install wording. Release notes.
v1.0.65 — reliability and MIDI drafts
Opening Izuki twice brings the existing window forward. Automatic update checks notify you; download and install from Settings → Updates when ready. Interrupted tasks cannot retain a success flag, and incomplete runs are not saved as successful flows.
Try “make a 32 bar trap arrangement at 100 bpm in A minor”. Izuki saves a new MIDI sketch with drums, bass, chords, melody and section markers, then reads it back and checks the notes, tempo and duration. Import it into a saved copy of your DAW project. From a paired phone, add “on my PC”. This does not analyze recordings, listen to a DAW, insert clips into a running project, or render finished audio. Release notes.
Say your wake word — like “Hey Nova”. A glowing liquid orb
appears, Izuki says “Mhm?”, and you just talk — like a voice call with a friend who
happens to be great with computers. It can:
- Answer questions about what's on your screen — “what does this error mean?”, “summarise this page”.
- Do things for you — “open Spotify and play something chill”, “reply to this email saying I'll be late”.
It looks, acts, looks again, and keeps going until the job is done.
- Point things out — “where's the export button?” and it circles it on your screen while it explains.
- Ask you when it's unsure — “Which account? Circle it for me.” Circle it, press Enter, and it carries on.
- Check with you before anything final — submitting, sending, buying or deleting always waits for your OK.
- Remember you — your name, your preferences, the habits it notices. You can see and delete every memory.
- Tap you on the shoulder — an important email, a meeting in 15 minutes, homework due tomorrow, a morning brief.
- Look things up for you — it searches the web and reads the pages itself, quietly, even sites you're signed in to.
- Do everyday things instantly — “scroll down”, “louder”, “pause”, “next song”, “go back”, “new tab”: done at once, no AI, like Siri.
- Work with any AI, even small free ones — they get a simple menu (Easy Mode) and Izuki finds the real button by name, so they act instead of explaining.
- Stop the instant you say so — Esc or Ctrl+Shift+Q, any time, in every mode.
Features
New in 1.0.14: repaired account routing, Mouse/Precision controls, three matching orb styles, recoverable flow cleanup, Discord delivery checks, and a video classroom. Release details and test steps.
| |
|---|
| 🎙️ Hands-free voice | Wake word runs on your PC (no audio leaves it until you talk to Izuki). Live transcript of your words, a voice-reactive orb, and you can talk over Izuki to interrupt it — like ChatGPT's voice mode. |
| 🗣️ A natural voice | Lifelike free neural voices that pause at commas and full stops like a person — no key needed. Or a voice that runs on your own PC, offline (Kokoro), Groq's expressive Orpheus (free key), Google's Gemini voice (free with your Gemini key), Azure Speech (the same Natural voices through Microsoft's official free plan), ElevenLabs (the most human voices, free monthly allowance), or ChatGPT's voice (needs a little OpenAI credit). |
| 🎭 30+ characters | Pick who Izuki is: warm Nova, deep calm Leo, British Sophie, Nigerian Ezinne, Naija Pidgin Chidi, Spanish Lucía, French, Swahili, Hindi, a hype coach, a butler, a pirate — or Rex, unfiltered and sarcastic (it swears and roasts you, only if you pick it). Rename it, pick any of 50+ voices, change speed and pitch, and give it your own personality. It talks that way everywhere — voice, chat, phone and Telegram. |
| 👀 Sees your screen | Understands screenshots and the real buttons/fields Windows reports, so clicks land on the right control. |
| 🔁 Look → act → look again | Multi-step tasks keep going across new windows, pages and pop-ups until the job is visibly done. |
| 🚀 Instant commands | "Scroll down", "louder", "quieter", "pause", "play", "next song", "go back", "refresh", "new tab", "close tab", "minimise this", "show desktop", "zoom in", "undo", "save" — done at once with no AI, so there's nothing to wait for and nothing to get wrong. Works by voice, typing, and from your phone. |
| 🎯 Safe Hands (accurate clicks) | Every click is checked in a few milliseconds before it happens: is the button really under the pointer (pages move as ads and pictures load — then it clicks where the button is now), is it greyed out, is another window on top of it (then it presses the button directly through Windows, or clears the cover first), and after typing it reads the box back and fixes a wrong letter. It clicks the visible part of a half-hidden link, places the pointer exactly on any monitor, and its own orb never catches the click. Say "click Subscribe" or "press Sign in" and it's clicked instantly, no AI needed. Built on Microsoft's UFO² research and Anthropic's computer-use guidance. |
| 🐞 |
Install
- Download
Izuki_…_x64-setup.exe from the latest release and run it.
Windows may say “Windows protected your PC” because the app is new and not yet code-signed —
click More info → Run anyway.
- The welcome tour opens. Get a free Gemini key at aistudio.google.com/apikey
(sign in → Create API key) and paste it into the tour. That's the only setup.
- Turn on hands-free in the tour and add a wake word — a free “Hey Nova” from
openwakeword.com (Draw → Hey Izuki → Wake words walks you through it).
Say it, and talk. No wake word? Hold Ctrl+Win+Space to talk any time.
The first time you use voice, Izuki downloads its speech models once (a few hundred MB), then works offline for listening and speaking.
Needs: Windows 10 or 11 (64-bit), a microphone (a headset works best), and internet for the AI brain.
macOS and Linux: coming soon — star the repo to hear when they land.
How to use it
| Do this | To |
|---|
| Say your wake word, e.g. “Hey Nova” | Start a conversation. It keeps listening between turns — say “that's all” to end it (or set how long it waits, up to 30 min). |
| Just talk while Izuki is answering | Interrupt it — it stops and listens. |
| Hold Ctrl+Win+Space | Push-to-talk without the wake word. |
| Hold Ctrl+Alt+Space, talk, let go | Dictate: what you said is typed where your cursor already is, in any app. Nothing is ever sent. |
| Hold Ctrl+D, draw, let go | Mark something and ask about it (Enter with nothing typed = “help me with this”). |
| Ctrl+Shift+Space | The full drawing overlay, for multi-mark jobs. |
| Esc (while Izuki is busy) or Ctrl+Shift+Q | Stop everything. |
| Say or type “quit Izuki” | Close the app completely. |
| “Remember that …” / “Forget …” | Manage what Izuki knows about you. |
Every shortcut can be changed in Settings → Shortcuts. To add a wake word: sign in at
openwakeword.com, find or train one (e.g. “Hey Nova”), download the ONNX file, and click
Add it under Draw → Hey Izuki → Wake words. (Models from that site are free for personal use.)
Privacy
- Wake-word detection and speech-to-text run on your PC. With Sharper hearing on (and a Gemini or Groq key), what you say to Izuki is also sent to that provider to transcribe — switch it off in Talk settings to keep speech on your PC.
- Phone messages go through Telegram to Izuki on your PC; only the one phone you paired is answered.
- When a request needs your screen, a screenshot is sent to the AI provider you chose — nowhere else. Plain conversation sends only text.
- API keys, memories and settings stay in
%APPDATA%\Izuki on your machine. Izuki never remembers passwords, keys or card numbers.
- No accounts, no telemetry.
What Izuki will not do on its own
Izuki drives your real mouse and keyboard, so these are enforced in the code, not
left to the AI's judgement — a free model is told to behave and sometimes doesn't.
- It won't type into a password box. It asks the box what it is first, and
refuses. It also never reads one, so a password can't reach the AI either.
- It won't act on the lock screen. A keystroke there goes into your PIN.
- It won't press a button that can't be taken back — Submit, Send, Pay,
Buy, Delete, Erase, Format, Share, Invite, Revoke, Close account and the rest —
unless you asked for that in so many words in that same request. Say "delete
that file" and it deletes it; say "sort this page out" and it does everything
else, then stops and asks.
- It won't pick or submit answers on a quiz or test. It will explain the
question and the idea behind it, and point at what matters, and you choose.
Ask it outright and it will tell you an answer — it just won't click it.
- It won't type a command into your terminal unless you spelled that command
out yourself.
- Esc stops everything, mid-task, mid-click, every mode.
Build from source
You need Node.js 20+ and Rust (plus the
Tauri prerequisites for Windows).
git clone https://github.com/nova-izuki/izuki.git
cd izuki
npm install
npm run tauri dev # run it
npm run tauri build # make the installer (src-tauri/target/release/bundle)
Built with Tauri 2, React and Rust.
For developers: Safe Hands 🖐️
Building your own AI that uses a Windows PC? Izuki's hands are a separate,
MIT-licensed Rust library: crates/safe-hands. It
gives your model the window's real buttons as a numbered list (so it answers
"click 7" instead of guessing pixels), checks every click before it happens
(really there? moved? greyed out? covered — then pressed directly), reads
typed text back, and places the pointer exactly on any monitor.
safe-hands = { git = "https://github.com/nova-izuki/izuki" }
For developers: voice-turns 🎙️
Building a voice assistant? packages/voice-turns
(TypeScript, MIT, zero dependencies) has two parts:
- EchoGate: tells the user talking over the assistant apart from the
assistant's own voice echoing into a laptop mic.
- QuietGate: lets an always-on wake-word detector rest in silence without
missing the first syllable.
Roadmap
- Voice personalities and characters, more voices and languages
- “Explain this lecture” mode: pause videos and annotate the screen while teaching
- Deeper help in creative apps (music, drawing, video) and games
- More ways to teach it your routines
Ideas and pull requests are welcome — open an issue.
☕ Support the builder
Izuki is free, and it's going to stay free. It's built by one person, late at night,
powered by a lot of coffee and a big dream.
If Izuki saved you some time, made you smile, or you just want to see where it goes next,
you can buy me a coffee:
💚 Tip on Cash App — $louismane2
Every tip — even a dollar — keeps the lights on and the next feature coming. And if you
can't tip, a ⭐ on this repo or telling a friend means just as much. Thank you. 🙏
About the builder
Hi, I'm Solomon Nwachukwu — an IT and cybersecurity student and the builder of Izuki.
Izuki is step one of something bigger. I'm working toward building real robot companions
(Nova Izuki is the name of that robot), an Afro-tech community where more of us
get to build the future instead of just using it, and one day a company big enough to
create real job opportunities for a lot of people. It's step by step — and Izuki is the first step.
Get in touch: [email protected] —
questions, ideas, collaborations or just to say hi.
Credits & licenses
Izuki is MIT-licensed (see LICENSE). It builds on:
Made with ❤️ by Solomon Nwachukwu.
