🎙️ Ari — Open-Source Windows AI Voice Assistant & Desktop Automation Agent

Voice control, desktop automation, local AI, memory, MCP tools, and an animated assistant — in one Windows app.

Ari is an open-source Windows AI voice assistant and autonomous desktop agent built with Python and PySide6. It combines wake-word voice control, speech-to-text (STT), text-to-speech (TTS), Windows automation, persistent memory, local or hosted LLMs, Model Context Protocol (MCP), plugins, and skills in one desktop application.

It is designed for people who want a Windows voice assistant that can do more than chat: Ari can route supported commands locally, hand complex goals to an agent workflow, use tools, verify results, remember useful context, and speak the result back.

[!NOTE]
Ari currently targets Windows 10/11 (64-bit). Optional direct local execution, Telegram remote commands, and remote embeddings are disabled by default.

What Ari can do

Area	Capabilities
Voice assistant	Wake word, voice shortcut, character click, Google STT, offline Whisper STT, streamed speech output
Windows automation	Supported local commands plus multi-step desktop tasks through tools and the autonomous agent
Fast local commands	Optional high-confidence path for commands such as time, running apps, screenshots, and volume control
Local AI	Ollama local LLMs, local CosyVoice3, local ONNX embeddings, and offline Whisper options
Hosted AI	OpenAI-compatible providers for models hosted locally or remotely
Memory	Relevant fact and conversation retrieval, explicit remember/forget commands, reviewable memory suggestions
Agent workflows	Plan, execute, verify, reflect, reuse strategies, interrupt active work, and resume supported flows
Extensions	Plugins, installable `SKILL.md` skills, MCP servers and tools
Remote control	Optional allow-listed Telegram commands through the same request pipeline
Languages	Korean, English, and Japanese UI/command routing

Why Ari

Windows-first, not browser-first. Ari is built around desktop interaction, voice control, system commands, and Windows automation.
Local-first options. Use Ollama, Whisper, CosyVoice3, and local embeddings when you want more of the stack to stay on-device.
Fast when a full model is unnecessary. Supported high-confidence commands can use the optional local decision path instead of waiting on an LLM.
More than a chatbot. Complex requests can enter a plan → execute → verify → reflect loop and use tools or reusable strategies.
A visible assistant. The desktop character reflects activity and mood, supports event-based speech, and can be clicked to start talking.
Built to extend. Add plugins, SKILL.md packages, MCP tools, or OpenAI-compatible model providers without replacing the core app.

Example interactions

Ari's exact behavior depends on enabled features and your selected model/provider, but supported requests include things like:

"What's the current time?"
"Set the volume to 30%."
"Take a screenshot."
"What apps are running?"
"Remember that I prefer concise answers."
"Forget what I told you about that preference."

For larger goals, Ari can pass the request into its autonomous agent workflow and use the tools available in your configuration.

Quick start

Requirements

Windows 10/11 (64-bit)
Python 3.11 for source installs
8 GB RAM recommended
4 GB GPU VRAM recommended when using local GPU-backed models

Install the Windows build

Download Ari-Setup-<version>.exe from GitHub Releases.

The installer uses Program Files\Ari by default. User settings, history, and runtime data are stored under %AppData%\Ari.

Run from source

git clone https://github.com/DO0OG/Ari-VoiceCommand.git
cd Ari-VoiceCommand\VoiceCommand
setup.bat
Ari.vbs

For local CosyVoice3, run setup.bat --with-tts. Use Ari.bat when you need startup diagnostics. See the Usage Guide for configuration, providers, voice settings, skills, and advanced features.

Local-first and privacy-aware defaults

Ari can use cloud services, but the project also supports a more local setup:

Ollama for local LLM inference
Offline Whisper for speech recognition
CosyVoice3 for local TTS
Local ONNX embeddings for memory and strategy retrieval
Remote embeddings disabled by default
Telegram integration disabled by default
Direct local command execution disabled by default
Plugins are not automatically loaded without user consent

The exact data flow depends on the providers and optional integrations you enable.

How it works

graph TD
    Wake["Wake word"] --> STT["STT: Google / Whisper"]
    Manual["Voice shortcut / character click"] --> STT
    Manual --> Warm
    Wake --> Warm["LLM connection prewarm"]
    STT --> Registry["Command registry"]
    Chat["Text chat / Telegram"] --> Handler["Request handler"]
    Registry --> Handler
    Handler --> Decision["Local decision engine"]
    Decision -- "Eligible + high confidence" --> Fast["Direct local command"]
    Decision -- "Uncertain / complex" --> Ack["Instant acknowledgement"]
    Decision -- "Uncertain / complex" --> LLM["LLM provider + tool calls"]
    Ack -.-> LLM
    Warm --> LLM
    LLM --> Tools["Tool execution"]
    Tools --> Policy["Tool-result follow-up policy"]
    Policy --> LLM
    Memory["Fact + conversation memory"] --> LLM
    Tools --> Agent["Autonomous agent"]
    Agent --> Loop["Plan / execute / verify / reflect"]
    Loop --> Strategy["Strategy memory / skills"]
    Strategy -.-> Agent
    Fast --> TTS["Sentence-streamed TTS"]
    LLM --> TTS
    Agent --> TTS
    Agent --> Character["Character mood / event speech"]
    TTS --> Character
    Stop["Stop / interrupt"] --> TTS
    Stop --> Agent

Verified local-decision evaluation

The optional local decision engine has a separate held-out evaluation for its supported command path:

7,407 generated evaluation examples
367 parser-confirmed direct selections
100.0% measured precision for those selections
0 false direct selections in that evaluation
Warm inference: p50 0.053 ms, p95 0.100 ms over 1,000 iterations on an AMD64 Windows desktop

These numbers evaluate the decision/parser path only. They do not measure microphone recognition accuracy, overall agent success rate, or every possible user request.

What's new in v1.1

Faster voice interaction with wake-time connection prewarming and instant acknowledgement
Wake word + command in one utterance
Faster end-of-speech handling and improved Whisper behavior
Streaming Edge TTS with caching and sentence-level playback
Interrupt Ari while it is speaking or generating a response
Explicit remember/forget commands and improved multilingual memory retrieval
Persistent mood, conversational context, and event-based character speech
Update checks, notifications, and installer reliability improvements
v1.2.0: voice cloning and emotional speech across TTS providers — new OpenAI-compatible TTS, ElevenLabs voice cloning and v3 emotion tags, and OpenAI custom voices
v1.2.0: fixes the installed app closing right after startup, and every release installer is now launched and checked before publishing
v1.2.1: Fish Audio and ElevenLabs TTS start playing from the first audio chunk, the TTS volume setting applies to every engine, and TTS failures are shown in the speech bubble
v1.2.1: speech bubble sits right above the character's head, "open Naver"-style requests fall back to the browser for known sites, and missing translations were filled in
v1.2.2: microphone auto-sensitivity is now off by default so quiet microphones are recognized, and fast local handling is on by default
v1.2.2: fixed Python tool runs failing with WinError 6, settings tabs that squashed their inputs, CosyVoice reloading after a stop, scheduled tasks and skills updates, and speech bubbles going off screen

See the v1.2.2 release notes.

For developers

Ari is a Python/PySide6 Windows desktop project with multiple extension points:

Useful areas for contributions include Windows automation, STT/TTS, local model integrations, PySide6 UX, plugins, skills, MCP workflows, reliability, and multilingual support.

Assets & credits

Default character illustrations by JAraTang: https://www.pixiv.net/users/78194943
DNFBitBitv2 font, official source: https://df.nexon.com/data/font/dnfbitbitv2

Check the font's usage terms before redistributing or reusing it outside this project.

License

Ari is released under the MIT License.