FatihMakes/Mark-LIV

Newest, latest and most advanced model that is able to control computers and even more. We're improving that model day by day, so follow up to know and see new models.

🎯 What is MARK LIV?

MARK LIV is a cross‑platform, voice‑first personal AI assistant that runs on Windows, macOS and Linux. It talks to you, listens to you, can see your screen or webcam, and can control the computer (open apps, adjust volume, move files, send messages, etc.). The standout visual element is a software‑rendered 3‑D human head that lip‑syncs, blinks, and moves its eyes in real time – all drawn with PyQt 6’s QPainter, so no GPU or extra graphics libraries are required.

The assistant is powered by Google Gemini 3.1 Flash Live (the “Gemini Live API”), which streams audio and text back‑and‑forth with ultra‑low latency. All of the assistant’s knowledge about itself (its name, OS, installed plugins, and even its limits) is generated at start‑up from the live system rather than being hard‑coded.

✨ Core Capabilities (as listed in the README)

Feature	What it does
Holographic Avatar	Animated human head in the HUD; real measured facial geometry, rendered entirely in software (no GPU).
Real Lip‑Sync	~50 mouth shapes per second derived from audio formants and the transcript, so the mouth forms the correct phonemes.
Language‑Free Mouth	One rule‑set works for Latin, Cyrillic, Greek and many other scripts; scripts without reliable spelling fallback to audio‑only shaping.
Facial Acting	Brows follow the phrase, eyes make saccades, natural blinking, nod on stressed syllables.
Face as Status Light	Looks away while thinking, meets eyes while listening, lids close when “asleep”, glances down at new content.
Push‑to‑Talk (Ctrl+Space)	Global hot‑key on Windows, window‑scoped elsewhere; microphone stays closed until the key is held.
Self‑Echo Guard	Detects and discards the assistant’s own voice from the microphone input, preventing it from answering itself.
Runtime Self‑Knowledge	Generates a prompt containing its name, OS, abilities and explicit limits each session; updates automatically when plugins are added/removed.
Wake Word	Local “Hey Jarvis” detection; sleeps when idle and never streams audio while asleep.
Instant Acknowledgment	Speaks a short, context‑aware reply in the user’s language the moment a longer task starts.
Unlimited Sessions	Sliding‑window context compression lets a single conversation run for hours without losing earlier context.
Plugin System	Drop a single `.py` file into `plugins/`; the assistant discovers it on next launch and can call its `run()` function.
Recallable / Persistent Memory	Facts are stored in `memory/long_term.json`; the UI shows a memory panel where you can view or delete entries.
Undo	Reverses file moves, renames, creations, writes, and setting changes performed by the assistant.
Real Confirmation	Irreversible actions (shutdown, restart, Wi‑Fi changes) require an explicit button press from the user.
Audio Device Picker	Lists only the microphones/speakers that the OS reports as functional; you can select by name.
Session Continuity	Changing devices, voice, or a dropped connection does not reset the conversation.
Live Theming & HUD	Re‑colour the whole HUD via a hue wheel or hex code; the avatar retints instantly.
Multi‑Mode Web Search	`news`, `research`, `price`, `compare`, `search` – Gemini‑grounded first, DuckDuckGo fallback.
System Control	Launch apps, adjust volume/brightness, toggle Wi‑Fi, power actions, etc., all by voice.
Visual Awareness	On‑demand screen capture and webcam frames are sent to Gemini and labelled.
Background Monitoring	Configurable topic watching (e.g., daily news) with natural‑language alerts.
Hardware Monitoring	Continuous CPU, RAM, GPU, temperature telemetry with spoken alerts.
Remote Dashboard	Pair a phone via QR code to control the assistant remotely.
Clipboard Intelligence	Copy text → floating panel offering translate, summarise, explain, fix, etc.
Assistant Customisation	Change name, voice, colour, etc., from the UI; changes take effect immediately.
Many other utilities	Flight finder, game updater, file processor, code helper, browser control, messaging (WhatsApp/Telegram), YouTube control, smart reminders, weather reports, etc.

📦 How to get it running (from the information in the README)

Clone the repository and install the listed Python dependencies (the project already uses PyQt6 and numpy; no extra graphics libraries are required). The README stresses that the app ships with zero extra dependencies beyond what it already needs.
Obtain a Gemini Live API key (required for the real‑time voice streaming). The assistant will not work without a valid key.
Run the main Python entry point (the README does not name the script, but typical projects use something like python main.py). The UI will appear with the HUD in the centre of the screen.
Configure audio devices via the built‑in picker if the default microphone or speakers are not correct.
Use the wake word (“Hey Jarvis”) or hold Ctrl+Space to start speaking. The assistant will respond with voice and animated facial cues.
Add plugins by dropping a .py file into the plugins/ folder; restart the app and the new skill will be available.
Explore the memory panel, theming options, and the various command categories (system control, web search, etc.) from the UI.

🛠️ Notable Technical Details (from the README)

Rendering – The avatar’s skull, neck, jaw, and rig are generated procedurally at start‑up; only a 25 KB facial mesh asset is stored on disk. All drawing is done with QPainter, guaranteeing identical appearance on any hardware, even headless VMs.
Lip‑Sync Algorithm – Combines a formant‑based audio analysis (20 ms slices) with the transcript to decide mouth shape, achieving language‑independent articulation.
Global Hot‑Key (Windows) – Implemented by polling virtual‑key codes 30 times per second; no extra library is used. macOS/Linux fall back to window‑scoped hot‑keys.
Self‑Echo Guard – Measures the audio device’s latency and subtracts the assistant’s own output from the microphone input, preventing accidental self‑responses.
Memory Architecture – Full facts are stored locally; only a compact “core” and an index of keys are injected into the Gemini prompt each turn. Missing facts are fetched on‑demand via a recall_memory tool.
Undo System – Tracks file system changes and setting modifications; a single UI button can roll them back.
No GPU Requirement – Because everything is software‑rendered, the assistant runs on machines without a dedicated graphics card.

⚠️ Limitations & Open Issues (as reported in the README)

Self‑echo guard works, but interrupting the assistant mid‑sentence is deliberately disabled in this release because it is highly hardware‑dependent.
Language support for scripts without phonetic spelling (Arabic, Chinese, Japanese, etc.) falls back to audio‑only shaping, which yields less detailed lip‑sync.
Global hot‑key works only on Windows; on macOS/Linux the push‑to‑talk key works only when the assistant window has focus.
Memory index size – If the memory grows beyond the prompt budget, the index may truncate, causing the assistant to claim it “doesn’t know” facts that are actually stored.
No GPU acceleration – While intentional, this means the avatar may look less sophisticated than a shader‑based 3‑D model on high‑end hardware.
Reliance on Gemini Live API – The assistant’s responsiveness and capabilities depend on the availability and quota of the Gemini service.

📚 Where to learn more

YouTube setup video – https://www.youtube.com/@FatihMakes (linked in the README).
Source code – The repository itself contains the prompt template (core/prompt.txt), the plugin architecture, and the rendering logic.
Gemini Live API documentation – external resource for obtaining an API key and understanding streaming limits.

In short: MARK LIV is a fully‑featured, voice‑driven AI assistant that adds a surprisingly expressive software‑rendered face as a status indicator. It is built on the Gemini Live model, works on all major desktop OSes without extra graphics dependencies, and is extensible via a simple Python plugin system. The README provides a thorough overview of features, design choices, and known quirks, making it a genuine, frontier‑technology project in the AI‑assistant space.