Files
Mark-LIV/readme.md
T
2026-06-06 17:41:38 +03:00

5.6 KiB

🤖 MARK XL — Local AI Assistant

J.A.R.V.I.S — Just A Rather Very Intelligent System
Cross-platform voice AI assistant running entirely on local hardware. No cloud APIs required.


Overview

MARK XL is a fully local, real-time voice and visual AI agent. It combines offline speech recognition, a locally hosted LLM (via Ollama), and text-to-speech to deliver a privacy-first personal assistant with OS-level control capabilities.

Successor to the previous mark, which used the Google Gemini Live API. MARK XL removes all cloud LLM dependencies while adding streaming responses, a dynamic configuration UI, and multi-language support.


Architecture

Microphone → STT (Whisper / Vosk)
                  ↓
           Ollama LLM (tool calling + streaming)
                  ↓
         Tool Execution (OS, Browser, Files …)
                  ↓
           TTS (EdgeTTS / Kokoro / ElevenLabs)
                  ↓
              Speaker

Core Components

Layer Technology Notes
STT faster-whisper / Vosk Fully offline. Auto language detection or forced locale.
LLM Ollama (any model) qwen2.5, llama3.2, mistral, etc. Streaming + tool calling.
TTS EdgeTTS / Kokoro / ElevenLabs EdgeTTS = free + internet. Kokoro = fully offline.
UI PyQt6 HUD overlay with system monitor, log panel, file drop zone.
Agent Custom task queue Multi-step planner + executor + error recovery.

Features

  • Streaming responses — TTS starts speaking on the first sentence, not after the full response
  • Tool calling — 18 built-in tools: browser control, file management, weather, YouTube, messaging, screen analysis, code helper, game updater, flight finder, and more
  • Long-term memory — Silently saves personal facts; recalled in every conversation
  • Live configuration — Change LLM model, STT engine, TTS voice without restarting (⚙ Configure button)
  • Ollama auto-start — Automatically launches ollama serve if it's not running
  • Model warmup — Pre-loads the LLM into memory during startup so the first message is as fast as subsequent ones
  • Multi-language STT — Set stt_language to auto (Whisper detects) or a specific locale (tr, de, fr, …)
  • Cross-platform — Windows, macOS, Linux (OS detected automatically at runtime)
  • File drop zone — Drag and drop images, PDFs, Word docs, CSV, audio, video for AI processing

Requirements

  • Python 3.11 or 3.12
  • Ollama installed and a model pulled (e.g. ollama pull qwen2.5:7b)
  • A microphone

Quick Start

# 1. Install Ollama → https://ollama.com
#    Then pull a model:
ollama pull qwen2.5:7b

# 2. Clone / download the project and launch
cd Mark-XL
python main.py

That's it. On first run MARK XL:

  1. Auto-installs base packages (PyQt6, numpy …) and restarts once
  2. Opens the Initialisation overlay — choose STT engine, LLM model, TTS engine
  3. Click INITIALISE SYSTEMS — engine packages install in the background (progress shown in log)
  4. JARVIS comes online

After setup, use the ⚙ CONFIGURE button in the right panel to change any setting at any time without restarting.


Configuration (config/api_keys.json)

{
    "stt_engine":         "whisper",
    "stt_model":          "base",
    "stt_language":       "auto",
    "llm_url":            "http://localhost:11434",
    "llm_model":          "qwen2.5:7b",
    "tts_engine":         "edgetts",
    "tts_voice":          "en-US-GuyNeural",
    "elevenlabs_api_key": ""
}
Key Values Default
stt_engine whisper / vosk whisper
stt_model tiny / base / small / medium / large-v3 base
stt_language auto or ISO code (tr, en, de …) auto
llm_url Ollama API base URL http://localhost:11434
llm_model Any model pulled in Ollama qwen2.5:7b
tts_engine edgetts / kokoro / elevenlabs edgetts
tts_voice Voice name / ID depending on engine en-US-GuyNeural

Built-in Tools

Tool Description
open_app Opens any application or website
web_search Web search and compare mode
weather_report Current weather for any city
send_message WhatsApp / Telegram messaging
reminder Timed reminders via Task Scheduler
youtube_video Play, summarize, trending videos
screen_process Screen capture + vision model analysis
computer_settings Volume, brightness, window management, shortcuts
browser_control Full Playwright browser automation
file_controller File/folder CRUD, search, disk usage
desktop_control Wallpaper, organize, clean desktop
code_helper Write, edit, explain, run code
dev_agent Build complete multi-file projects
agent_task Multi-step autonomous task execution
computer_control Direct mouse/keyboard control
game_updater Steam / Epic Games install & update
flight_finder Google Flights search
file_processor Process images, PDFs, CSV, audio, video

Keyboard Shortcuts

Key Action
F4 Mute / unmute microphone
F11 Toggle fullscreen

TTS Engine Comparison

Engine Internet Quality Cost
EdgeTTS Required Good Free
Kokoro No Excellent Free (local model ~100 MB)
ElevenLabs Required Best Paid API

License

MIT — FatihMakes Industries