Repository navigation
Releases
Releases · PRN-6/PRIVACY68
PRIVACY68 v1.6.5
What's Changed in 1.6.5
- Gmail Automation Controller: Direct web compose, inbox navigation, and speech-to-email drafting with multi-account
authuserrouting support. - Interactive Click-to-Configure Plugin Drawers: Click any plugin card in the Control Center to expand an inline configuration drawer with live input boxes (e.g. Gmail address, WhatsApp contact, browser homepage) and 1-click test actions.
- Zero-Latency Natural Language Email Parsing: Instant parameter extraction (
to,subject,body) directly without LLM latency. - Unified IPC Plugin Settings: Dashboard bridge supports
set_plugin_setting,get_plugin_setting, andexecute_plugin_actionwith profile persistence. - Universal AI Core (Cloud APIs & Local LLMs): Full support for Google Gemini (2.5 Flash), Groq LPU (GPT-OSS 120B / Llama 3.3 70B), OpenRouter, Claude, OpenAI, and local Ollama / LM Studio with custom Base URLs.
- 100% Click-to-Speak Hardware Isolation: Zero background mic recording; microphone hardware is opened strictly on-demand.
- 3-Tier Security Guard: Real-time permission gating (Level 1 Safe instant execution, Level 2 Sensitive notification, Level 3 Critical confirmation prompt with 15s timeout).
- Verified Stability: Full test suite passing 107/107 automated tests (100% pass rate).
Requirements & Acceleration
- Universal AI: Works immediately out-of-the-box with free Cloud API keys (Gemini, Groq, OpenRouter) or local offline inference (Ollama / LM Studio).
- GPU Acceleration (Optional): Runs on CPU out of the box, or click 'Download CUDA' in the Control Center Hardware card for NVIDIA GPU acceleration.
PRIVACY68 v1.6.4
What's Changed in 1.6.4
- Gmail Automation Controller: Direct web compose, inbox navigation, and speech-to-email drafting with multi-account
authuserrouting support. - Interactive Click-to-Configure Plugin Drawers: Click any plugin card in the Control Center to expand an inline configuration drawer with live input boxes (e.g. Gmail address, WhatsApp contact, browser homepage) and 1-click test actions.
- Zero-Latency Natural Language Email Parsing: Instant parameter extraction (
to,subject,body) directly without LLM latency. - Unified IPC Plugin Settings: Dashboard bridge supports
set_plugin_setting,get_plugin_setting, andexecute_plugin_actionwith profile persistence. - Universal AI Core (Cloud APIs & Local LLMs): Full support for Google Gemini (2.5 Flash), Groq LPU (GPT-OSS 120B / Llama 3.3 70B), OpenRouter, Claude, OpenAI, and local Ollama / LM Studio with custom Base URLs.
- 100% Click-to-Speak Hardware Isolation: Zero background mic recording; microphone hardware is opened strictly on-demand.
- 3-Tier Security Guard: Real-time permission gating (Level 1 Safe instant execution, Level 2 Sensitive notification, Level 3 Critical confirmation prompt with 15s timeout).
- Verified Stability: Full test suite passing 107/107 automated tests (100% pass rate).
Requirements & Acceleration
- Universal AI: Works immediately out-of-the-box with free Cloud API keys (Gemini, Groq, OpenRouter) or local offline inference (Ollama / LM Studio).
- GPU Acceleration (Optional): Runs on CPU out of the box, or click 'Download CUDA' in the Control Center Hardware card for NVIDIA GPU acceleration.
PRIVACY68 v1.6.3
What's Changed in 1.6.3
- Gmail Automation Controller: Direct web compose, inbox navigation, and speech-to-email drafting with multi-account
authuserrouting support. - Interactive Click-to-Configure Plugin Drawers: Click any plugin card in the Control Center to expand an inline configuration drawer with live input boxes (e.g. Gmail address, WhatsApp contact, browser homepage) and 1-click test actions.
- Zero-Latency Natural Language Email Parsing: Instant parameter extraction (
to,subject,body) directly without LLM latency. - Unified IPC Plugin Settings: Dashboard bridge supports
set_plugin_setting,get_plugin_setting, andexecute_plugin_actionwith profile persistence. - Universal AI Core (Cloud APIs & Local LLMs): Full support for Google Gemini (2.5 Flash), Groq LPU (GPT-OSS 120B / Llama 3.3 70B), OpenRouter, Claude, OpenAI, and local Ollama / LM Studio with custom Base URLs.
- 100% Click-to-Speak Hardware Isolation: Zero background mic recording; microphone hardware is opened strictly on-demand.
- 3-Tier Security Guard: Real-time permission gating (Level 1 Safe instant execution, Level 2 Sensitive notification, Level 3 Critical confirmation prompt with 15s timeout).
- Verified Stability: Full test suite passing 107/107 automated tests (100% pass rate).
Requirements & Acceleration
- Universal AI: Works immediately out-of-the-box with free Cloud API keys (Gemini, Groq, OpenRouter) or local offline inference (Ollama / LM Studio).
- GPU Acceleration (Optional): Runs on CPU out of the box, or click 'Download CUDA' in the Control Center Hardware card for NVIDIA GPU acceleration.
PRIVACY68 v1.6.2
What's Changed in 1.6.2
- Universal AI Core (Cloud APIs & Local LLMs): Full support for Google Gemini (2.5 Flash), Groq LPU (GPT-OSS 120B / Llama 3.3 70B), OpenRouter, Claude, OpenAI, and local Ollama / LM Studio with custom Base URLs.
- 100% Click-to-Speak Hardware Isolation: Zero background mic recording or circular buffer idle capture; hardware is opened strictly on-demand.
- Acoustic Peak Normalization: Speech waveforms are scaled dynamically via
norm = (audio / peak) * 0.9for studio-grade transcription across both quiet and loud environments. - 3-Tier Security Guard: Real-time permission gating (Level 1 Safe instant execution, Level 2 Sensitive notification, Level 3 Critical confirmation prompt with 15s timeout).
- Control Center v1.6.2: Dark glassmorphic dashboard with 1-click model selection chips, custom Base URLs, and real-time connectivity testing.
- Autonomous Skill Crystallization: Instant fast-path execution of learned desktop automation tasks with self-healing code generation.
- Verified Stability: Full test suite passing 87/87 automated tests.
Requirements & Acceleration
- Universal AI: Works immediately out-of-the-box with free Cloud API keys (Gemini, Groq, OpenRouter) or local offline inference (Ollama / LM Studio).
- GPU Acceleration (Optional): Runs on CPU out of the box, or click 'Download CUDA' in the Control Center Hardware card for NVIDIA GPU acceleration.
PRIVACY68 v1.6.1
What's Changed in 1.6.1
- Clean Minimalist Control Center: Removed emojis and cluttered icons for an ultra-sleek, professional desktop management interface.
- Multi-Provider Cloud AI Slots: Simultaneous slots for Google Gemini (Free 1.5/2.0 Flash), Groq Cloud (Free Llama 3.3 70B), Anthropic Claude, OpenAI, and Ollama.
- Autonomous Skill Crystallization: 0ms fast-path execution of learned desktop automation tasks.
- Mobile Web Remote Server: Control the assistant from your phone or tablet on local Wi-Fi (Port 8765).
- 3-Tier Security Guard: Strict permission gating (Safe Level 1, Sensitive Level 2 with prompt, Critical Level 3 blocked).
- 100% Free Speech & Push-to-Talk Privacy: Zero background mic recording; Google Web Speech ASR on-demand.
Requirements
- Ollama: Local Ollama service is recommended for executing agentic OS automation commands (or configure an API key).
- GPU Acceleration (Optional): Can run on CPU out of the box, or click 'Download CUDA' in Hardware settings for NVIDIA GPU acceleration.
PRIVACY68 v1.5.0
What's Changed in v1.5.0
- Ollama Detection & Management: Real-time probe in Control Center displaying local/cloud Ollama service status and one-click controls.
- On-Demand CUDA 12 Runtime: One-click download and extraction of cuBLAS/cuDNN DLLs directly from the dashboard for instant GPU acceleration on compatible NVIDIA GPUs.
- Dynamic Path Resolution: Enhanced Agent Brain path resolution across OneDrive redirected folders (Desktop, Documents, Downloads).
- Alexa Wake Word: Fast and accurate wake phrase detection using Faster-Whisper and Silero neural VAD.
- Streamlined Footprint: Removed heavy legacy biometric models for a faster and cleaner desktop installation.
Requirements
- Ollama: Local Ollama service is recommended for executing agentic OS automation commands (or configure an API key).
- GPU Acceleration (Optional): Can run on CPU out of the box, or click 'Download CUDA' in Hardware settings for NVIDIA GPU acceleration.
PRIVACY68 v1.4.3
What's changed
- Fixes the frozen Windows build so LiteLLM can load the
cl100k_basetokenizer encoding. - Routes common Chrome and WhatsApp launch requests directly to their built-in plugins.
- Preserves the Agent Brain's actual error details in command results.
- Fixes frozen dashboard launching and isolated Agent Brain subprocess execution.
Requirements
- Model-backed commands require the local Ollama service and the configured model; they are not bundled.
- Chrome and WhatsApp launch commands require the corresponding application to be installed.
The project uses LiteLLM with Ollama; it does not use Open Interpreter.
PRIVACY68 v1.4.2
PRIVACY68 v1.4.2: Windows installer and GitHub release downloads
PRIVACY68 v1.4.1
📋 PRIVACY68 Changelog
All notable changes to PRIVACY68 are documented here.
Versioning follows Semantic Versioning — MAJOR.MINOR.PATCH.
[v1.4.1] — 2026-10-21
✨ New Features & Improvements
🪟 App-Context Command Routing (computer_use/app_focus.py, actions/fast_lane.py)
- Active App Window Context Awareness: Fast Lane commands automatically resolve context and target the active foreground application (e.g. Chrome, WhatsApp, VS Code, Notepad).
- Targeted Application Actions: Context-aware routing for actions like closing tabs, navigating, search, and message sending directly mapped to the currently active application.
💬 Enhanced WhatsApp Desktop Automation (extra_plugins/whatsapp_plugin.py, plugins/base_plugin.py)
- Robust Message Automation: Extended WhatsApp Desktop plugin with improved contact resolution, UI navigation, and message dispatch.
- Enhanced Plugin Base & Manager: Improved plugin management lifecycle and hook callbacks.
🔗 Compound Command Planner (actions/planner.py)
- Multi-Command Parsing: Seamlessly plan and execute multi-intent composite voice utterances sequentially.
[v1.4.0] — 2026-10-19
✨ New Features & Ecosystem Upgrades
🔌 Dynamic Plugin Architecture & Hot-Reload (plugins/manager.py, ui/dashboard_bridge.py)
- Runtime Plugin Hot-Reload: Install, enable, disable, and reload plugins dynamically without restarting the assistant.
- Robust External Plugin Resolution: Added
sys.pathandsys.modulesaliasing with MRO-basedBasePluginresolution to seamlessly support third-party and custom marketplace plugins. - Enhanced Built-In & Marketplace Plugins: Updated plugins for Google Chrome, Brave, WhatsApp Desktop, GitHub Manager, Notepad, VS Code, and PowerPoint with rich keyboard and window automation.
🛠️ Antigravity IDE Automation & Developer Tools (computer_use/dev_tools.py, actions/executor.py)
- Automated IDE DevTools: Intelligent window detection, foreground elevation, and native terminal execution for Antigravity IDE and VS Code.
- Voice-Triggered Git Workflow: Dedicated voice automation to launch terminal sessions, stage, commit, and push changes to GitHub directly from voice command (
"push to GitHub").
🎙️ Dynamic Wake Word Engine & Profile Reload (speech/streamer.py, plugins/profile_manager.py)
- Dynamic Wake Word Switching: Switch active wake words (
Jarvis,Privacy68,Alexa,Nova,Friday, etc.) dynamically via dashboard configuration without restarting the application. - Phonetic Alias Expansion: Added comprehensive phonetic alias matching for Jarvis and voice commands to eliminate transcription variances in noisy environments.
[v1.3.0] — 2026-10-18
✨ New Features & Architecture Upgrades
⚡ 3-Lane Command Architecture (actions/router.py, actions/fast_lane.py, actions/api_lane.py)
- Fast Lane (Deterministic): 100% local, zero-latency execution for all known system commands, application automations, and media controls without contacting an LLM.
- API Lane (Cloud Reasoning): Fast cloud intelligence using OpenAI or Groq when complex questions, coding tasks, or open-ended reasoning are requested.
- Local LLM Lane (Private Fallback): Full offline privacy via local Ollama models with automatic fallback.
🎯 ML Intent Classification & Parameter Extraction (actions/intent_classifier.py, actions/param_extractor.py)
- Dedicated machine-learning intent classification pipeline with unified command mapping.
- Dedicated parameter extractor trained to 100% accuracy on standard system entities (applications, URLs, paths, numbers).
- Drastically streamlined Fast Lane codebase (-640 lines of brittle regex) in favor of high-speed vector models.
🎙️ Robust Wake Word & Speech-to-Text (speech/streamer.py, config.py)
- Phonetic Alias Matching: Broad phonetic coverage for "Privacy 68", "Privacy sixty eight", "Privacy six eight", "Privacy 6 8", and digit-separated variations.
- 2-Gram Multi-Word Fuzzy Matching: Accommodates natural speaking cadence, varied accents, and different microphone responses.
- Audio Pre-processing Upgrades: Audio peak normalization, lowered idle-scan RMS threshold (from 0.018 to 0.005), and relaxed VAD gating ensure quiet or low-gain microphones trigger reliably.
- Whisper Decoder Priming: Expanded hotwords and initial prompts to prevent Whisper from dropping numeric/alphabetic variations.
📱 Mobile Web Remote Server (server/remote_server.py)
- Built-in lightweight local web remote allowing mobile devices on the same Wi-Fi network to control PRIVACY68 with QR code pairing.
[v1.2.0] — 2026-10-12
✨ New Features
☁️ Cloud LLM API Support (actions/router.py)
- OpenAI, Groq, and Custom API providers are now first-class citizens alongside local Ollama.
- The
LLMRouterauto-selects the routing path based on the user's savedllm_providerandapi_keysettings. - Cloud route implements an OpenAI-compatible REST client (no SDK dependency) with a hard 8-second timeout and graceful Ollama fallback if the cloud call fails.
- Supported cloud models out of the box:
gpt-4o-mini(OpenAI),llama-3.3-70b-versatile(Groq), and any custom OpenAI-compatible endpoint.
🎙️ Enhanced Voice Detection & Speaker Verification (speech/voice_auth.py)
- Upgraded ECAPA-TDNN voice biometrics model with improved Kaldi 80-channel filterbank.
- Tightened energy-based
trim_speechpre-processing to strip silence before embedding extraction. - Verification threshold tuning: owner scores
0.85–0.95; rejection fires at< 0.45. - Dynamic disk-profile reload — profile changes in the dashboard take effect immediately without restart.
📱 WhatsApp Contact Manager (ui/dashboard/index.html, ui/dashboard_bridge.py)
- New Contact Manager panel in the Dashboard for the WhatsApp plugin.
- Add named contacts with display name, phone number, and spoken aliases.
- Fuzzy matching (
_resolve_contact) now maps natural speech (e.g., "send message to mom") to the correct WhatsApp chat name with high accuracy.
🔌 LLM & API Settings in Dashboard
- New AI Provider selector in Settings: Local Ollama / OpenAI / Groq / Custom.
- API Key, Base URL, and cloud model fields — with a live Test Connection button.
- API key is stored locally in
user_profile.json(never transmitted unless you use cloud providers).
🐛 Bug Fixes
- Fixed
LLMRouter.reload()not respecting disabled plugins when building the tool name map. - Fixed
profile_manager.pyrace condition on concurrent settings saves from the dashboard. - Voice enrollment via the Dashboard now shares the exact same audio path as live verification.
🧪 Tests Added
tests/test_cloud_llm_router.py— unit tests for OpenAI & Groq cloud routing paths.tests/test_voice_auth_unit.py— unit tests for ECAPA-TDNN embedding extraction and cosine similarity verification.
📦 Build & Distribution
build.batupgraded to a full two-step release pipeline: PyInstaller EXE then Inno Setup installer.- Inno Setup step is auto-skipped gracefully if ISCC.exe is not installed.
build_exe.pynow carriesAPP_VERSION = "1.2.0"and auto-includesversion_info.txtfor Windows EXE metadata if present.installer.issupdated to outputPRIVACY68-Setup-v1.2.0.exe.
[v1.1.0] — 2026-10-05
✨ New Features
- Silero VAD neural voice activity detection replaces RMS energy thresholding.
- Speaker Recognition & Voice Biometrics (Owner-Only Voice Lock) with ECAPA-TDNN.
- Two-step conversational wake word pipeline: "Alexa" triggers 7-second active listening window.
- WhatsApp launch mechanism with UWP -> Standalone EXE -> Protocol URI -> Web fallback chain.
- Mobile Web Remote: WebSocket server for controlling PRIVACY68 from a phone browser.
- NVIDIA CUDA runtime on-demand downloader (no bundled DLLs required in slim build).
- Control Center Dashboard (WebView2) with plugin management, voice lock, and diagnostics.
📦 Build
- PyInstaller
.specandbuild_exe.pywith optional--bundle-cudaflag. - Inno Setup
installer.issfor a one-click Windows installer.
[v1.0.0] — 2026-09-20
- Initial public release.
- Faster-Whisper ASR, wake word detection, System Tray, and Floating HUD overlay.
- Plugin system with Chrome, WhatsApp, VS Code, and PowerPoint plugins.
- Fast Lane deterministic Windows automation (< 5 ms, zero LLM overhead).
- Ollama integration for AI intent routing (qwen2.5:0.5b).
PRIVACY68 v1.4.0
📋 PRIVACY68 Changelog
All notable changes to PRIVACY68 are documented here.
Versioning follows Semantic Versioning — MAJOR.MINOR.PATCH.
[v1.4.0] — 2026-10-19
✨ New Features & Ecosystem Upgrades
🔌 Dynamic Plugin Architecture & Hot-Reload (plugins/manager.py, ui/dashboard_bridge.py)
- Runtime Plugin Hot-Reload: Install, enable, disable, and reload plugins dynamically without restarting the assistant.
- Robust External Plugin Resolution: Added
sys.pathandsys.modulesaliasing with MRO-basedBasePluginresolution to seamlessly support third-party and custom marketplace plugins. - Enhanced Built-In & Marketplace Plugins: Updated plugins for Google Chrome, Brave, WhatsApp Desktop, GitHub Manager, Notepad, VS Code, and PowerPoint with rich keyboard and window automation.
🛠️ Antigravity IDE Automation & Developer Tools (computer_use/dev_tools.py, actions/executor.py)
- Automated IDE DevTools: Intelligent window detection, foreground elevation, and native terminal execution for Antigravity IDE and VS Code.
- Voice-Triggered Git Workflow: Dedicated voice automation to launch terminal sessions, stage, commit, and push changes to GitHub directly from voice command (
"push to GitHub").
🎙️ Dynamic Wake Word Engine & Profile Reload (speech/streamer.py, plugins/profile_manager.py)
- Dynamic Wake Word Switching: Switch active wake words (
Jarvis,Privacy68,Alexa,Nova,Friday, etc.) dynamically via dashboard configuration without restarting the application. - Phonetic Alias Expansion: Added comprehensive phonetic alias matching for Jarvis and voice commands to eliminate transcription variances in noisy environments.
[v1.3.0] — 2026-10-18
✨ New Features & Architecture Upgrades
⚡ 3-Lane Command Architecture (actions/router.py, actions/fast_lane.py, actions/api_lane.py)
- Fast Lane (Deterministic): 100% local, zero-latency execution for all known system commands, application automations, and media controls without contacting an LLM.
- API Lane (Cloud Reasoning): Fast cloud intelligence using OpenAI or Groq when complex questions, coding tasks, or open-ended reasoning are requested.
- Local LLM Lane (Private Fallback): Full offline privacy via local Ollama models with automatic fallback.
🎯 ML Intent Classification & Parameter Extraction (actions/intent_classifier.py, actions/param_extractor.py)
- Dedicated machine-learning intent classification pipeline with unified command mapping.
- Dedicated parameter extractor trained to 100% accuracy on standard system entities (applications, URLs, paths, numbers).
- Drastically streamlined Fast Lane codebase (-640 lines of brittle regex) in favor of high-speed vector models.
🎙️ Robust Wake Word & Speech-to-Text (speech/streamer.py, config.py)
- Phonetic Alias Matching: Broad phonetic coverage for "Privacy 68", "Privacy sixty eight", "Privacy six eight", "Privacy 6 8", and digit-separated variations.
- 2-Gram Multi-Word Fuzzy Matching: Accommodates natural speaking cadence, varied accents, and different microphone responses.
- Audio Pre-processing Upgrades: Audio peak normalization, lowered idle-scan RMS threshold (from 0.018 to 0.005), and relaxed VAD gating ensure quiet or low-gain microphones trigger reliably.
- Whisper Decoder Priming: Expanded hotwords and initial prompts to prevent Whisper from dropping numeric/alphabetic variations.
📱 Mobile Web Remote Server (server/remote_server.py)
- Built-in lightweight local web remote allowing mobile devices on the same Wi-Fi network to control PRIVACY68 with QR code pairing.
[v1.2.0] — 2026-10-12
✨ New Features
☁️ Cloud LLM API Support (actions/router.py)
- OpenAI, Groq, and Custom API providers are now first-class citizens alongside local Ollama.
- The
LLMRouterauto-selects the routing path based on the user's savedllm_providerandapi_keysettings. - Cloud route implements an OpenAI-compatible REST client (no SDK dependency) with a hard 8-second timeout and graceful Ollama fallback if the cloud call fails.
- Supported cloud models out of the box:
gpt-4o-mini(OpenAI),llama-3.3-70b-versatile(Groq), and any custom OpenAI-compatible endpoint.
🎙️ Enhanced Voice Detection & Speaker Verification (speech/voice_auth.py)
- Upgraded ECAPA-TDNN voice biometrics model with improved Kaldi 80-channel filterbank.
- Tightened energy-based
trim_speechpre-processing to strip silence before embedding extraction. - Verification threshold tuning: owner scores
0.85–0.95; rejection fires at< 0.45. - Dynamic disk-profile reload — profile changes in the dashboard take effect immediately without restart.
📱 WhatsApp Contact Manager (ui/dashboard/index.html, ui/dashboard_bridge.py)
- New Contact Manager panel in the Dashboard for the WhatsApp plugin.
- Add named contacts with display name, phone number, and spoken aliases.
- Fuzzy matching (
_resolve_contact) now maps natural speech (e.g., "send message to mom") to the correct WhatsApp chat name with high accuracy.
🔌 LLM & API Settings in Dashboard
- New AI Provider selector in Settings: Local Ollama / OpenAI / Groq / Custom.
- API Key, Base URL, and cloud model fields — with a live Test Connection button.
- API key is stored locally in
user_profile.json(never transmitted unless you use cloud providers).
🐛 Bug Fixes
- Fixed
LLMRouter.reload()not respecting disabled plugins when building the tool name map. - Fixed
profile_manager.pyrace condition on concurrent settings saves from the dashboard. - Voice enrollment via the Dashboard now shares the exact same audio path as live verification.
🧪 Tests Added
tests/test_cloud_llm_router.py— unit tests for OpenAI & Groq cloud routing paths.tests/test_voice_auth_unit.py— unit tests for ECAPA-TDNN embedding extraction and cosine similarity verification.
📦 Build & Distribution
build.batupgraded to a full two-step release pipeline: PyInstaller EXE then Inno Setup installer.- Inno Setup step is auto-skipped gracefully if ISCC.exe is not installed.
build_exe.pynow carriesAPP_VERSION = "1.2.0"and auto-includesversion_info.txtfor Windows EXE metadata if present.installer.issupdated to outputPRIVACY68-Setup-v1.2.0.exe.
[v1.1.0] — 2026-10-05
✨ New Features
- Silero VAD neural voice activity detection replaces RMS energy thresholding.
- Speaker Recognition & Voice Biometrics (Owner-Only Voice Lock) with ECAPA-TDNN.
- Two-step conversational wake word pipeline: "Alexa" triggers 7-second active listening window.
- WhatsApp launch mechanism with UWP -> Standalone EXE -> Protocol URI -> Web fallback chain.
- Mobile Web Remote: WebSocket server for controlling PRIVACY68 from a phone browser.
- NVIDIA CUDA runtime on-demand downloader (no bundled DLLs required in slim build).
- Control Center Dashboard (WebView2) with plugin management, voice lock, and diagnostics.
📦 Build
- PyInstaller
.specandbuild_exe.pywith optional--bundle-cudaflag. - Inno Setup
installer.issfor a one-click Windows installer.
[v1.0.0] — 2026-09-20
- Initial public release.
- Faster-Whisper ASR, wake word detection, System Tray, and Floating HUD overlay.
- Plugin system with Chrome, WhatsApp, VS Code, and PowerPoint plugins.
- Fast Lane deterministic Windows automation (< 5 ms, zero LLM overhead).
- Ollama integration for AI intent routing (qwen2.5:0.5b).