JARVIS
NEXUS
Current Build — v1
A fully local AI OS orchestrator. Modular, decoupled, and engineered for near-zero latency — dual-model cognitive processing, thread-safe multi-layer memory, 30 built-in tools, and 100% offline by design.
What's New in v1
Dual-Model cognitive brain with TTFT KV-Cache pre-ignition at boot.
Producer-Consumer TTS architecture — zero silence gaps between sentences.
Smart Interruption with real-time mic buffer flushing and fuzzy matching.
Thread-safe multi-layer SQLite memory: Working, Episodic, Semantic, Procedural.
Zombie Process Management — 100% LLM uptime with self-healing recovery.
Setup Wizard webview GUI + autostart sync for Windows and Linux.
Production-First Engine
The Orchestrator Engine
A State Machine and Agentic loop managing memory, language modes, and permissions. Internal Command Processor with Fuzzy Matching for instant execution — zero LLM overhead for known commands.
Dual-Model Brain
Switches dynamically between normal_model (fast tasks) and overthink_model (deep logic). Immutable/Dynamic prompt layers, native JSON tool calling, and Zombie Process Management keep it alive.
Multi-Layer Cognition
Thread-safe SQLite with Mutex locks routing to Working, Episodic, Semantic, and Procedural memory. One consolidated auto-classification API — sentence-transformers for offline semantic vector search.
Zero-Latency Sensory
faster-whisper STT with dynamic hardware scaling (float16/int8) and real-time interrupt flush. Piper TTS double-queue Producer-Consumer architecture eliminates every silence gap.
OS & Web Orchestration
curl_cffi browser impersonation for headless deep research with Jina AI fallback. Strict OS workspace boundaries — sensitive paths require explicit permission gates.
Decoupled WatchDog
A completely decoupled background thread that syncs RAM tasks, monitors states, and fires proactive voice reminders with snooze logic — without touching the main engine loop.
30+ Tools, Zero Setup
Every tool is registered at boot and invoked via natural speech. No commands to memorize, just say what you want. The LLM picks the right tool and chains them automatically.
search memoryMemoryDigs through your personal long-term database for stored facts, preferences, and past conversations.
"When is my brother's birthday?"
save to memoryMemoryStores any fact, preference, or event permanently in your local memory. You own every byte.
"Save my passport number"
search webWebSilently fetches real-time web data for JARVIS to analyze and summarize — invisible to the browser.
"What is the current gold price in Egypt?"
run scenario the impossible tool
Every other tool has a fixed purpose. This one has no ceiling. Want JARVIS to boot a dev environment, kill background apps, switch your audio profile, and open your playlist in one command? Write the script once. Name the file. Done, forever.
How it works
Copy one of the prompts below
Send it to any AI with your task description filled in
Save the generated script to your JARVIS Run Directory
Name the file whatever you'd say to call it
"Jarvis, run [your script name] scenario", that's it, forever
// Example
File: work mode.ps1
Voice: "Jarvis, run work mode scenario"
Result: VS Code opens, Spotify pauses, Do Not Disturb activates.
AI Prompt Templates
Fill in your task, paste to any AI, get a ready-to-run script back. No coding required.
Scripts run with JARVIS's permissions. Review before placing in your Run Directory.
Persona Studio
JARVIS ships with two built-in personas (Classic and FRIDAY). But you can build any character you want a cold analyst, a chaotic gremlin, a hype man. Copy the template below, fill in the fields, and send it to any AI to get a ready-to-paste persona prompt.
Built-in Personas
Polished British executive assistant. Witty, loyal, mildly sarcastic. Addresses you as Sir or Boss.
Strictly professional, robotically cold, zero humor. Declarative sentences only. Pure execution.
Anything you imagine. Use the template on the right to generate and paste your own character.
Build Your Own
Copy this template, fill in the blanks, paste it into any AI assistant. It will generate a persona prompt you can add directly inside JARVIS Settings.
The AI is instructed to keep output under 580 chars and never embed real names matching JARVIS's internal persona format.
Engine Requirements
What JARVIS NEXUS needs to run. The Setup Wizard handles detection and download automatically on first launch.
Ollama Framework
Local AI Orchestrator
The foundational local inference engine. JARVIS manages it automatically — TTFT KV-Cache pre-ignition, self-healing model builds, and zombie process recovery.
Download from OllamaQwen Models / Any GGUF
Powered by Ollama
Drop any .gguf into models/llm/ — JARVIS auto-detects and builds the Modelfile. Native JSON tool calling preferred; falls back to Regex parsing if needed.
Faster-Whisper
by OpenAI · CTranslate2 backend
Dynamically scales compute (float16, int8) based on your hardware. Strict VAD filtering and real-time interrupt handling at the buffer level.
Piper TTS
Jarvis Voice Package
Ultra-fast local synthesis with pre-compiled Regex normalization — code blocks and math symbols stripped before vocalization. Producer-Consumer queue for gapless output.
Download from HuggingFaceRoadmap
Where JARVIS Nexus is heading. Every planned item is open for contributions.