A personal AI desktop assistant for Windows — chat, voice, and a huge set of built-in agents for everyday tasks, backed by a local LLM (via Ollama) with optional Google Gemini fallback for when you want more power.
Built by Venom.
This repository ships the packaged Windows application only. Source code is not published here.
Grab the latest build from the Releases page → download the .zip → extract it anywhere → run Marvell_AI.exe from inside the extracted folder.
No installer, no admin rights needed — it's a portable app. On first run, Windows SmartScreen may warn about an unrecognized publisher (normal for apps not code-signed by a commercial certificate) — click "More info" → "Run anyway."
Marvell is an always-on assistant that lives in your system tray and understands both typed and spoken commands.
| 🗣️ Chat & Voice | Type or talk to it; wake-word and push-to-talk support; barge-in (talk over a reply to interrupt it), continuous conversation for natural back-and-forth, and automatic mic calibration to your room. |
| 🌐 Agent Network | A live map of every specialist agent around a central core that reacts to your voice in real time, with per-agent status and real usage history. |
| 🧠 Local-first AI | Runs on a local Ollama model by default (private, no internet needed for basic chat), with an in-app model downloader and an optional Gemini fallback for tougher questions. |
| 💬 Messaging | WhatsApp, Telegram, Slack, Discord — read, reply, and automate. |
| ✅ Productivity | Reminders, calendar, email, notes, habit tracking, journaling, recurring automation rules. |
| 📁 Files & media | File creation, image generation, video/audio tools, screen recording. |
| 👨💻 Dev tools | Code generation (with a dedicated coding model + copy-paste-ready code blocks), git helpers, a sandboxed scripting console. |
| 🖱️ Computer use | Can look at your screen and carry out a task via mouse/keyboard, narrating each step with a hard step limit for safety. |
| 🏠 & more | Smart home, shopping/price tracking, market prices, focus/Pomodoro mode with distraction-site blocking, and dozens more agents. |
| 💰 Finance | Budgets, subscriptions, savings goals, invoices, loan/mortgage calculator, net worth, investment & crypto portfolio tracking. |
| 🛡️ Security & wellness | Breach monitoring, screen-lock automation, secure file shredder, sleep/mood tracking, guided breathing exercises, document expiry reminders. |
| 🎓 Learning | Language tutor with spaced-repetition quizzes, study plans, debate practice, daily trivia. |
| 📊 Dashboards | Usage/cost analytics, model response-time benchmarking, wellness trend correlation. |
Voice
- Real-time voice reaction — the Agent Network core now responds to your actual voice as you speak, driven by live microphone level rather than a canned animation.
- Barge-in — start talking while Marvell is replying and it stops immediately, instead of talking over you.
- Continuous conversation — it stays listening briefly after a reply, so a follow-up needs no wake word.
- Automatic mic calibration — measures your room's noise floor on mic-on and tunes the speech threshold to it, so it stops cutting you off early in a noisy room.
- Offline speech recognition (optional) — switch
stt_enginetoofflineto recognise speech locally via faster-whisper: no network, no quota, and nothing leaves your machine. Falls back to the online engine automatically if no local model is installed. - Louder replies, and a fix for emoji in a reply silently dropping the voice to a quieter fallback.
Ad skipping
- YouTube ads are now actually skipped, automatically. Previously this only announced itself and fired a blind keypress. Videos opened from Marvell now play in its own browser, where it can find and click the real skip button — and run out ads that aren't skippable.
Fixes
- Marvell no longer starts with Windows unless you ask it to. An entry left behind by an older build was launching it at every login regardless of the "Launch when Windows starts" checkbox; that entry is now cleaned up and the checkbox genuinely controls it.
- Dry-run mode — preview what an action would do before it actually happens.
- Guest mode — hard-blocks anything that sends/writes/controls something, for letting someone else use it safely.
- Opt-in clipboard history — off by default, since it can capture sensitive text.
- Local API — an optional local REST endpoint (off by default, token-protected) for scripting Marvell from other tools, plus a browser extension for sending page content or a right-click selection straight to it.
- Windows 10/11 (64-bit)
- For local AI: Ollama installed, with at least one model pulled — Marvell will detect this and help you download a model on first run
- A microphone, if you want to use voice features
Marvell is free to use. If it's saved you time and you'd like to support development, donations are welcome:
USDT (Binance Pay ID): 92555348
Not required, always appreciated. 🙏
See LICENSE for usage terms.
Found a bug or have a feature request? Open an issue.