- JavaScript 81.3%
- HTML 8.1%
- Python 4.5%
- Batchfile 3.4%
- Shell 2.7%
213d9c7b71
Decision: she needs about 7.4 GB of models and a GPU - a 4.9 GB language model, a 2 GB image checkpoint, her voice and her transcriber - so she runs where the hardware is, and a phone was only ever a window onto that. Removed mobile/ and android/ entirely. A real render at S23 size had already shown the layout was not mobile-ready anyway: the control bar ran off the right edge, the status line was clipped mid-sentence and the orb sat off-centre from horizontal overflow. Keeping the phone route meant fixing that first; dropping it makes the question moot. Swept every dangling reference out of the docs and both launchers. The LAN address is still printed, reworded from "your phone, same wifi, tap Install" to "same network" - reaching her from another machine still works and costs nothing. The PWA manifest and service worker stay: they are valid on desktop and the app installs from a desktop browser too. Windows and Linux now each have their own setup guide, and the main README points at both. Verified after stripping: the product boots, serves every app file, and reports all four capabilities. |
||
|---|---|---|
| app | ||
| codec | ||
| data | ||
| image-server | ||
| linux | ||
| models | ||
| serve | ||
| server | ||
| tokenizer | ||
| ui | ||
| voice-server | ||
| windows | ||
| .gitignore | ||
| local-tools-prompt.txt | ||
| Modelfile | ||
| package.json | ||
| README.md | ||
| Unity1.txt | ||
Unity 18+ — local
Uncensored Unity, running entirely on this machine. She talks, she speaks
out loud, she hears you, and she draws — local text, local image
generation on your GPU, local voice. Her persona is read straight off the disk
from Unity1.txt on every single message. No cloud, no account, no container,
no filter. The only network calls in the whole app go to 127.0.0.1.
Run it
| Your machine | Do this | Setup guide |
|---|---|---|
| Windows | double-click windows\start.bat |
windows/README.md |
| Linux | bash linux/start.sh |
linux/README.md |
⛔ This is a PC application. She needs about 7.4 GB of models and a GPU — a 4.9 GB language model, a 2 GB image checkpoint, and her voice and transcriber on top. That is a desktop workload, so she runs where the hardware is.
That is the whole thing. The first run sets itself up. It starts Ollama, builds her model, and downloads her image model (~2 GB) and her voice (~550 MB) in the background, then opens her at http://localhost:8790.
Chat works immediately. Pictures switch on when the image setup finishes, and her voice switches on when the voice setup finishes — the page notices on its own and starts using them, with no restart.
Talking to her
Type, or press talk and speak. She answers out loud in her own voice, and the orb is her actual waveform — the string spikes with her volume and shifts colour with her pitch, from your microphone as well as from her.
Ask her for a picture and she draws it on your GPU. Whatever you describe wins: name an outfit, or no outfit, or cover her in something, and that is what you get — her look stays hers, everything else is yours to set.
| Button | What |
|---|---|
| talk | Hold a conversation out loud. She hears you and answers |
| mute mic | Stop her hearing you without stopping her talking |
| mute voice | Read her instead of hearing her |
| attach | Give her a picture — a file, a paste, or a drag onto the window |
| gallery | Everything she has drawn. View, Save as…, or delete each one |
Pictures are kept as coefficient fields, not files, so the gallery costs kilobytes and every picture re-renders at whatever size your screen wants. Save as… writes a full-size PNG wherever you choose.
Editing her
Unity1.txt is Unity — all of her, sent verbatim on every single message.
Change a line, hit send, she is already different. No rebuild, no restart.
local-tools-prompt.txt is the only other thing she is told: how to make a
picture locally. Separate file, so her own words are never touched by tooling.
Folders
| Folder | What |
|---|---|
windows\ |
start.bat (everything) + the image and voice installers — setup guide |
linux/ |
start.sh (everything), stop.sh + the same installers — setup guide |
app/ |
Her app — the orb, chat, gallery, viewer, voice |
serve/ |
The app's server |
codec/ |
The wavelet codec every picture and every second of audio passes through |
server/ |
The text and image backend that holds her persona |
image-server/ |
Local Stable Diffusion (safety checker OFF) |
voice-server/ |
Her voice and her ears |
models/ |
Her image checkpoint and voice land here. Drop any .safetensors in and it is used |
ui/ |
The backend's own simple page |
What is running
| Port | What |
|---|---|
| 8790 | Her app — this is the one you open |
| 4545 | The backend: her persona, chat and image routing |
| 11434 | Ollama — local text |
| 7860 | Local Stable Diffusion — local images |
| 7861 | Her voice and her ears |
Check the app at once: http://localhost:8790/api/state — it reports what she can currently do, and anything still installing simply reads as not yet available rather than as broken.
Swapping her look
Any .safetensors checkpoint dropped into models\ is picked up on the next
start (first by name). The one it installs is Realistic Vision V5.1 —
photorealistic and uncensored. Civitai has thousands more.
Notes
Uncensored by design, local by construction. Nothing you say to her, and nothing she draws for you, leaves this machine.