Pati

A desktop companion that lives on your screen, talks with a local AI, remembers you, and gets things done. Think of the old desktop assistants, rebuilt for today: no ads, no tracking, and the AI runs on your own computer.

Türkçe

• Lives on your desktop. A small character above every window. It stands on the taskbar and on the top edges of your windows, walks off them and falls; drag it anywhere or throw it across the screen. It wanders around and falls asleep when you step away.
• Chats with a local AI. Connects to llama.cpp, Ollama or LM Studio running on your machine. Each character has its own personality.
• Remembers you. It keeps a brain: a folder of plain Markdown with your daily chats, a summary of each day and the facts you shared. Open it in Obsidian and edit what it knows.
• Gets things done. "Open Spotify", "turn the volume down", "take a note: …", "how's my battery?". It asks before acting, and you can say "always".
• Works with your files. "Create list.txt on the desktop: milk, bread", "add cheese to the shopping list", "what's in Downloads?", "move the report into the homework folder", "undo". Desktop, Documents and Downloads only, text files only (txt, md, csv); it asks first and never overwrites, deletes or touches programs and scripts.
• Talks by voice. Press the shortcut (Ctrl+Shift+Space by default) and speak; whisper.cpp turns it into text on your computer, and you see and hear the answer.
• Speaks up when it matters. A short briefing when you sit down in the morning, a nudge after 90 minutes without a break, a word when it is very late.
• Looks at your screen when you ask. "What's on my screen?", "explain this error" or Look at my screen in the menu: after asking, it takes one screenshot and shows it to a vision model (llama.cpp with --mmproj, or qwen2.5vl in Ollama). The image is never saved.
• Focus mode. "Focus" (or "focus for 50 minutes") starts a pomodoro: it stays quiet, then calls a 5-minute break, and a long one every fourth round.
• Reminders. "Every 60 minutes, drink water", "every day at 17:00, study", or just tell it in chat: "remind me in 20 minutes".
• Make it yours. Add your own character with a folder of PNG frames and one JSON file. Turkish and English, modern or classic look.

Built-in characters: Tungarix, a rock wolf in sunglasses, and Atlas, a calm holographic assistant.

Optional fan packs (extra characters) can be installed in Settings > Characters > Fan packs. Nothing is fetched until you press "Load list"; the packs come from a separate repository (tungarix/pati-fan-packs), not from this one. They are unofficial fan drawings.

Install

Download Pati Setup x.y.z.exe from Releases, or build it yourself (below). The installer is not code-signed, so Windows SmartScreen asks "Run anyway".

Pati lives in the system tray: right-click the tray icon for settings, characters and quit.

Do this	What happens
Click the character	Menu: chat, joke, time, tip, settings, quiet for an hour, hide
Drag	Move it; it remembers where
Double-click	It jumps
Right-click	Settings, character, hide, quit

Connect a local AI

Pati talks to any OpenAI-compatible server on your computer. It looks on the usual ports by itself (8080, 8081, 11434, 1234); you can also list servers in Settings > AI, in order of preference: put a smart, big model on top and a small, fast one below it. Pati uses the first one that is running.

• llama.cpp: llama-server -m model.gguf --jinja. --jinja enables tools. If you start it with --api-key, Pati asks for the key in the chat balloon the first time.
• Ollama: install it and ollama pull a model; Pati finds it on port 11434.
• LM Studio: load a model and start the local server.

Recommended model: Pati-4B

Pati-4B is a small model fine-tuned for Pati: setting reminders, remembering what you tell it, opening apps and sites, controlling music, taking notes, and short, warm small talk, mostly in Turkish. It is Qwen3.5-4B (Apache-2.0) with a LoRA trained on about 1,200 example conversations. It needs about 5 GB of graphics memory (it also runs on the processor, slowly). On Pati's own test set (158 cases) it scores 82%, against 67% for the first Pati-4B; details in docs/PATI-4B.md. It is made for chat and small tasks, not for long answers, code or facts you need to rely on.

Pati offers to download it on first start (about 3.4 GB, with llama.cpp), or use Settings > AI > Pati-4B > Install. It downloads with progress, checks the files, starts the model when Pati starts and stops it when Pati quits. Files go to %LOCALAPPDATA%\Pati\model; Remove in the same place deletes them.

Advanced / do it yourself: download llama.cpp (CUDA build for an NVIDIA card, Vulkan for others) and, from the pati-4b-v3 release, the two model files Pati-4B-v3-Q4_K_M-0000{1,2}-of-00002.gguf plus mmproj-4B-F16.gguf (from the pati-4b-v1 release; the vision file did not change) into one folder, then run llama-server -m Pati-4B-v3-Q4_K_M-00001-of-00002.gguf --mmproj mmproj-4B-F16.gguf --jinja -ngl 99 --port 8080. Pati finds it on port 8080 by itself.

Talk by voice

Your voice never leaves your computer: whisper.cpp turns it into text locally. Download whisper-cli.exe from its releases and put a multilingual model in a models folder next to it (ggml-small.bin is a good start; models ending in .en only know English). Pati finds whisper-cli.exe on PATH and the model in a models folder beside or above it; if it can't, pick both in Settings > Talking by voice, where you can also change the shortcut. Press, speak; Pati stops listening when you pause (or press the shortcut again).

Better voice (Piper)

The Turkish neural voice comes built in: the installer includes Piper and the tr_TR-dfki-medium voice, so nothing to download. In English, Pati speaks with the Windows voice (the built-in voice is Turkish only). You can still point Settings > Speech at your own piper.exe and voice (any voice from rhasspy/piper-voices); a voice you pick is used in every language. If Piper fails, the Windows voice speaks. Piper has no pitch setting.

Credits: Piper (MIT), espeak-ng inside it (GPL-3.0, source; Piper runs as a separate program), and the voice from DFKI (CC BY-NC-SA 4.0, non-commercial). Details in LICENSES.txt next to the bundled Piper.

A model of about 8-9B parameters is a good start on an 8 GB GPU. Pati works around the habits of small models: it handles plain commands itself instead of trusting the model to call a tool, and checks numbers (battery, time) instead of letting the model guess.

The brain

By default Documents\Pati Beyni, changeable in settings:

• Chats/2026-10-17.md: one note per day of chat. Pati writes a short summary at the top and reads the last few to pick up where you left off.
• Memory.md: what Pati knows about you, one line per fact. Add a line and it learns; delete one and it forgets.
• Notes.md: notes you asked it to take.
• Feedback.jsonl: answers you rated 👍 or 👎 in the chat, each with the whole conversation behind it. A few hundred of these are enough to fine-tune a small model in your character's style.

Open the folder as a vault in Obsidian once; after that Settings > Memory > Open in Obsidian takes you there.

Privacy

• The AI runs on your computer; chats, memory and notes stay in local files.
• Tools that reach outside the app are off until you turn them on: reading the clipboard, and looking things up online (weather from Open-Meteo, Wikipedia).
• Actions on the computer (opening apps, websites, files and folders, writing and moving files, music and volume) ask first. Only http(s) addresses are ever opened.
• The only thing Pati does on its own online is look for a new version on GitHub (a few hours apart, only in the installed app). You can turn it off in Settings > Updates; it always asks before installing.
• No telemetry, no accounts, no ads.

Your own character

Settings > Character > Open folder opens %APPDATA%\pati\characters. Put a folder there for your character and press Reload. The folder name is the character's id.

characters/
  my-cat/
    manifest.json
    character.json
    stand.png  walk1.png  walk2.png  sleep.png

manifest.json:

{ "schema": 2, "type": "character", "name": "My Cat", "version": "1.0.0", "author": "You" }

character.json:

{
  "language": "en",
  "scale": 4,
  "accent": "#f59e0b",
  "animations": {
    "idle":  { "frames": ["stand.png"], "fps": 1 },
    "walk":  { "frames": ["walk1.png", "walk2.png"], "fps": 4 },
    "talk":  { "frames": ["talk.png", "stand.png"], "fps": 6 },
    "sleep": { "frames": ["sleep.png"], "fps": 1 },
    "drag":  { "frames": ["held.png"], "fps": 1 },
    "fall":  { "frames": ["fall.png"], "fps": 1 },
    "dust":  { "frames": ["dust1.png", "dust2.png"], "fps": 6 }
  },
  "voice": { "lang": "en-US", "rate": 1, "pitch": 1.2 },
  "persona": "You are My Cat, a sleepy but curious cat who lives on the user's desktop.",
  "lines": {
    "greet": ["Meow! I'm here."],
    "menu": "What shall we do?",
    "jokes": ["..."],
    "tips": ["..."],
    "wake": ["I'm awake!"],
    "sleep": "Nap time…",
    "thanks": "Thanks",
    "hush": "OK, I'll be quiet.",
    "lateNight": ["It's late. Bed?"],
    "breakTime": ["Time for a short break."]
  }
}

• Only animations.idle (and a name) is required. fall plays in the air (else drag), dust once after a landing, guitar now and then when idle. Give walk a stride (one step's length in frame pixels) and its frames follow the distance walked, so the feet do not slide. A missing animation falls back to idle; missing lines fall back to defaults.
• lines can be one set in language, or one per language: { "tr": {...}, "en": {...} }. persona works the same way; it is the system prompt that gives the character its voice in chat.
• scale enlarges pixel art with crisp edges; leave it at 1 for drawn images. Keep the character under about 150 px wide.
• accent colours the modern balloon.
• A pack with the same id as a built-in one replaces it.

Tungarix's frames are cut from one drawn sprite sheet (art/tungarix-sheet.webp) by scripts/slice-tungarix.py: it removes the background, finds the poses and puts every frame on the same canvas with the feet on one line, so it works for any sheet laid out the same way. Atlas is drawn with geometry by scripts/make-atlas.py.

Development

npm install
npm start

Command	What it does
npm start	Builds and runs the app
npm test	Unit tests (Node's built-in runner)
npm run typecheck	TypeScript check
npm run dist	Builds the Windows installer into release/

Layout: src/main (windows, tray, AI, brain, tools, reminders, settings), src/core (behaviour state machine), src/preload.ts (the only bridge between windows and the main process), src/renderer (pet and settings windows), characters/ (built-in packs), locales/ (Turkish and English).

Releasing

Update version in package.json, run npm run dist (it first downloads Piper and the Turkish voice into build/piper/ once, hash-checked, which makes the installer about 85 MB bigger), and attach the installer, latest.yml and the .exe.blockmap from release/ to the GitHub Release (tag v<version>); without latest.yml installed copies won't see the update. Four-part hotfix versions (built with -c.buildVersion=0.4.4.1) keep the three-part package.json version in latest.yml, so updaters won't see them: bump the patch number when an auto-update is wanted.

Plans: docs/ROADMAP.md. Changes: CHANGELOG.md.

License

MIT. See LICENSE.