Talk to your PC. It listens, answers, and types for you.
Hold a button, speak, release — your words appear at the cursor in whatever app is focused. Or open a voice conversation that can actually do things: search the web, open your apps, run tasks on your machine.
No subscription. Bring your own OpenAI key. From $0.0045 per minute.
Read this in: English · Español · Français · Deutsch · Português (BR) · 简体中文 · 日本語
Windows 10 or 11. One line in PowerShell is the whole install:
irm https://raw.githubusercontent.com/RizN91/jarvis/main/install.ps1 | iexIt finds your Python, clones Jarvis into %LOCALAPPDATA%\Jarvis\app, builds the
virtual environment, installs the pinned dependencies, runs a smoke check, and
adds a Start Menu shortcut.
It is per-user and reversible: it never elevates, never installs Python for you,
and never changes your execution policy, your Defender settings or your global
Python. -DryRun prints exactly what it would do and changes nothing;
-Uninstall removes it again, and will not touch your data folder without a
typed confirmation.
The first launch opens a setup wizard — microphone, speakers, shortcuts, wake word and budget — and asks for your OpenAI API key, which goes straight into Windows Credential Manager and never into a file.
Nothing is sent anywhere until you finish the wizard and start dictating, apart from the connection tests you explicitly choose to run.
Or install by hand
git clone https://github.com/RizN91/jarvis.git
cd jarvis
setup.cmd
run.cmdOr straight from the source tree:
py -3.11 -m venv .venv
.venv\Scripts\python.exe -m pip install -r requirements.txt
.venv\Scripts\python.exe -m jarvismacOS and Linux are not supported yet. install.sh says so rather than
pretending, and Platform support explains why.
You never have to click anything.
|
Say “Hey Jarvis” Turn the wake word on and just start talking. It is detected on your machine by a small local model, so while it is listening for the phrase no audio leaves your computer at all. “Hey GPT” works too, and you can set your own phrase. |
Or hold a button The default is your mouse's side button (XButton2): hold it, speak, release — the words are already at your cursor. Prefer the keyboard? Tap F8. |
| You press | What happens |
|---|---|
| Mouse side button (hold) | Dictate into whichever app has focus |
F8 |
Start and stop dictation, hands-free |
Ctrl + Alt + Space |
Open a voice conversation |
Ctrl + Alt + Pause |
Emergency stop — ends the session and any running tool |
Esc |
Cancel whatever is happening right now |
Every one of them is rebindable in the setup wizard — to a different key, a different mouse button, or nothing at all.
A real session, not a mockup. The pill and the text arriving in Notepad are the shipped code running — the clip is generated from the app itself, so it cannot quietly drift from what you download.
▶ Watch the 88-second demo with sound · 1080p, 12 MB
You think faster than you type. You also repeat yourself all day — the same email shape, the same commit message, the same summarise-this-document request.
Dictation tools solve half of that and are either a subscription, or a toy, or they silently press Enter and send your half-finished message.
Jarvis does both halves, costs pennies, and stays out of the way. It has no subscription, no account, no telemetry, and no server of its own.
|
Say it, and it's typed
…lands in your email, your editor, your ticket, wherever the cursor is. |
Or ask for something
…it searches, then answers out loud, in a conversation you can talk over. |
The pill in all six states — idle, dictating, working, listening, speaking, error. It never steals focus, and it idles when nothing is happening.
Twelve-step onboarding: it checks your microphone, speakers, shortcuts and key before you rely on it.
Every screen has a light theme, and the whole UI is translated — see below.
You bring your own OpenAI key, so you pay OpenAI directly — no markup, no subscription, no seat licence. Two engines, and the choice is yours:
| Engine | Model | Cost | Feels like |
|---|---|---|---|
| Economy (default for dictation) | gpt-transcribe |
$0.0045 / min | about 27 hours per dollar; the most accurate for verbatim text |
| Live | gpt-live-1 |
$0.05 / min | about 20 minutes per dollar; conversational, handles barge-in and turns |
For scale: twenty minutes of dictation a day, five days a week, costs about 45 cents a month on Economy. The assistant is the expensive part, and only while you are actually talking to it.
Jarvis also meters its own spend and stops at the daily and monthly ceilings you set. Those ceilings are Jarvis's own counters — it cannot read your OpenAI account budget, and it never claims to.
Dictation
- Types into any app — editors, browsers, chat, terminals, ticketing systems.
- Never presses Enter, so it cannot send a half-written message.
- Refuses password fields; holds text for review in terminals rather than auto-typing into a shell.
- Your own vocabulary — names, jargon, product names — passed to the transcriber as hints, which measurably improves accuracy.
- Undo that will not eat your own typing.
Voice assistant
Ctrl+Alt+Spaceopens a conversation with real barge-in — talk over it and it stops.- It can search the web and act on your machine through a tool layer where every action needs your explicit Yes, with the exact arguments shown.
- Local coding agents (Codex, Claude Code) can be handed long jobs, and Jarvis reports which one it used.
Always
- Local wake words — "Hey Jarvis", "Hey GPT". Runs on your machine
(
sherpa-onnx); no audio leaves the device while it is listening for the wake phrase. - The pill. A per-pixel-alpha floating overlay that shows what it heard and what it is doing, and is genuinely click-through and non-activating.
- Privacy by construction — no telemetry, no cloud database, no analytics, no account. Raw audio is not stored.
- Light and dark themes. Reduced-motion support.
The entire setup and settings UI ships in 15 languages, with a switcher in Settings:
English · Español · Français · Deutsch · Italiano · Nederlands ·
Polski · Português (BR) · Русский · Türkçe · العربية (RTL) ·
हिन्दी · 中文 (简体) · 日本語 · 한국어
Your language is detected from your system on first run. Missing a language? It is one file and a pull request — see CONTRIBUTING.md.
Windows 10 and 11 are fully supported. The core is built on Win32 — global
input hooks, SendInput text insertion, layered windows for the pill, DPAPI for
the key, WTS for lock/suspend detection. That is what makes it feel native, and
it is also why it is not yet portable.
macOS and Linux are not supported yet. Running there exits with a clear message rather than a traceback. A port is on the roadmap; it is a real project, not a flag.
What is coming next, roughly in order:
- More models, not just OpenAI — Claude and Gemini as the assistant's brain, and local models via Ollama for a fully offline mode.
- Better voices and a voice picker, including your own voice profile.
- Themes and skins — more than light/dark, and a pill that matches your desktop.
- True streaming dictation — today Live records then replays, so release-to-result includes the length of what you said.
- macOS port (Accessibility API + Quartz) and Linux (X11).
- Plugins — let people add their own spoken commands and tools.
- A signed installer, so SmartScreen stops warning.
Ideas, votes and complaints are welcome in Issues — the roadmap follows what people actually ask for.
A short version; the real detail is in ARCHITECTURE.md.
- A tiny tray process owns global hooks, audio and the pill. It idles at 2 processes, ~53 MB and ~0.3 % of one core.
- The settings window is a separate process, so WebView2 (~440 MB) exists only while you have it open.
- Speech goes straight to OpenAI over a WebSocket. There is no Jarvis server and no localhost port — nothing to attack, nothing to go down.
- Your data (history, vocabulary, spend) is SQLite on your own disk. Nothing is synced anywhere.
- Your API key is stored only in Windows Credential Manager, DPAPI-wrapped, and never written to a config file, database, log, command line or the UI.
- A redaction filter scrubs key-shaped strings from every log record.
- No telemetry. No analytics. No account.
- Raw audio is not written to disk.
- Startup at login is opt-in and reversible; uninstall removes everything.
See SECURITY.md for how to report a vulnerability.
Bug reports, translations, themes and new tool verbs are all welcome — start with CONTRIBUTING.md. It explains the test suites, how to add a language, and the handful of rules that have already prevented real bugs (no secrets in logs, never synthesise Enter, never let dictated text become instructions).
The test suite is ~486 assertions across 13 free suites, and it types into real applications rather than mocks. There is a lot to build if you want to help.
MIT — do what you like with it.
Dependency and model licences are listed in THIRD-PARTY-NOTICES.md, including one honest caveat about the wake-word model's terms that is worth reading before any commercial use.
Jarvis is not affiliated with, endorsed by, or sponsored by OpenAI. You supply your own API key and are responsible for your usage.
If Jarvis saves you time, a ⭐ helps other people find it.