When you call it
It wakes up when you say its name. And you can turn listening off.
Early access
You say it. It does it.
A companion that lives on your Windows PC and runs it with you, by voice or by text. It thinks with its own model, on your graphics card, and by default sends nothing of yours out.
- GitHubavailable now
- Steamsoon
- itch.iosoon
Windows · dedicated 4 GB graphics card · English and Spanish
BAXY
Idle
Spotify opened and started playing Kupla's Lavender, which was already a lo-fi track.
Open app · SpotifyPlay music · lo-fi
Open Spotify and play some lo-fi to study
Trailer · 72 s
Watch it work on a real desktop.
Five requests recorded on a Windows PC, with anything personal blurred. What you see, BAXY did.
With sound. BAXY's voice is synthetic.
Watch it on YouTube
What it does today
Ask the way you'd ask a person.
Pick a request. The screenshot and the reply come from the real take, word for word.
BAXY replies
Spotify opened and started playing Kupla's Lavender, which was already a lo-fi track.
Open app · SpotifyPlay music · lo-fi
- Programs and windowsOpens, closes and arranges what's on your screen.
- Inside your appsWorks through the windows of whatever you have open: Settings, File Explorer, Calculator, Paint, Microsoft Store, Spotify, Discord, Steam…
- Chained requests"Open the calculator, work out 125 plus 75 and paste it into Notepad."
- The webSearches the internet and sums up what it finds.
- Windows settingsVolume, brightness, Bluetooth and others.
- Its voiceWakes up when you say "BAXY" and talks back with its own synthetic voice. Without a microphone, it works by text.
How it behaves
It acts, then it tells you.
It's still learning and sometimes gets things wrong. So it tells you what it did without being asked, in one sentence.
When it doesn't understand
It asks one short question instead of guessing.
When it works
It tells you in a sentence: "I set the volume to 40%." The detail is one click away.
When something fails
It tells you what failed. And if it can't do something, it says so instead of improvising.
Yours and private
By default, what's yours stays on your PC.
Runs on your PC
- The language modelBAXY's own fine-tune of Qwen3.5-4B
- Its hearingSpeech recognition and the wake word
- Its voiceSynthetic, with its own character
- Its memoryYou can see, edit and delete all of it
Goes out only if you ask
What you ask of the web
It can search the internet when needed, and the window shows "Went online".
"Another model", if you turn it on
It's off by default. If you connect your own key from an AI company, what you give BAXY goes to that company, and your account pays for it. It's in the 1.3.0 installer on GitHub.
And you're in charge.
Anything destructive it asks first: deleting, buying, sending something to someone else or closing unsaved work. Everything else it just does, then tells you. If you'd rather it didn't ask, you turn that off yourself, on purpose.
Requirements
It needs a dedicated 4 GB graphics card.
Check this before you download. Here's what we know today, card by card.
Tested
NVIDIA, 4 GB or more
Uses CUDA. Driver 551.61 or later.
In testing
AMD, 4 GB or more
The installer accepts it and uses Vulkan. On an RX 7600, 9 of the 12 planned cases went well, by text. Still to do: 3 cases, the window and the voice.
Not tested
Dedicated Intel
It would use Vulkan, but it hasn't been tried on a real card.
Not supported
Integrated graphics or CPU only
BAXY may not work, or may take very long to start.
| System | 64-bit Windows 11. The installer also accepts Windows 10, with less testing. |
|---|---|
| Disk | About 11.1 GB free while installing. Afterwards it takes 5.89 GB with NVIDIA (4.70 GB without), plus about 1 GB of cache. |
| Memory | 16 GB of RAM recommended. Still to be confirmed on a clean machine. |
| Internet | To install. After that it works offline, except for what you ask of the web. |
| To talk to it | A microphone and speakers. Without them, it works by text. |
| Languages | English and Spanish. Interface and voice in both. Mix the two and it still understands you. |
Where to get it
Get BAXY.
Free on GitHub today. Coming soon to Steam and itch.io at USD 9.90. The GitHub installer isn't a cut-down version, and BAXY is in early access everywhere.
GitHub
FreeBAXY 1.3.0
It checks your PC, downloads what it needs (5.19 GB with NVIDIA, 4.53 GB without), verifies each part and leaves BAXY in the Start menu.
It isn't code-signed, so SmartScreen may warn you. Choose "More info", then "Run anyway".
Download the installer See the project on GitHub See the release and its notesSteam
USD 9.90CLP 9,990
Steam installs and updates it for you, with the model, voice and speech recognition included.
Not on sale yet. The link will appear here.
Coming soon to Steamitch.io
USD 9.90CLP 9,990
With the itch app, updates download only what changed. The price is the minimum: you can pay more if you'd like to support it.
Not on sale yet. The link will appear here.
Coming soon to itch.ioWhy it's sold in stores
I build BAXY on my own, and so far I've done it with my own time and money. That no longer stretches far enough to give it what it needs. The code is free and will stay free. If BAXY is useful to you as it is today, buying it is the most direct way to let me keep working on it.
Questions
Worth knowing first.
Is it finished?
No, and I don't want to pretend it is. It's usable every day, but it has bugs: sometimes it can't complete an action, and it tells you so ("I couldn't confirm it"). I'm building it on my own, and I want people with different PCs, voices and habits to use it.
When does it leave early access?
I won't set a date I'm not sure I can keep, and I'm not promising specific features. It leaves early access when what it does, it does every time on any PC that meets the requirements.
What hasn't been tested yet?
- Version 1.3.0 on a real AMD or Intel card: an AMD card was simulated for the installer.
- A freshly formatted Windows without the Visual C++ runtime or WebView2. The installer installs both.
- A real microphone and speakers on this version: the test used recorded sentences.
- The first start: while BAXY checks its files, the window may say "The model isn't available". It then starts by itself.
- There's no echo cancellation: while it speaks, the wake word may hear it.
Can I use another model?
Yes, in Settings → Model. It's optional and off by default. OpenAI and Google haven't been tested with a real key, and other models do worse than BAXY's own, which BAXY is tuned for.
What's the license?
BAXY's own code is MIT. Each third-party component keeps its own license: the voice (Supertonic 3) is under OpenRAIL-M, which forbids some uses, and the avatar's videos and images aren't covered by the MIT license. The full list is in the third-party notices.
How do I report a bug?
On GitHub, with the exact phrase you said and what it did. Every bug reported that way becomes a test case.
