Invisible personal copilot. Always listening. Never shows up.

Jards is a local macOS agent that continuously listens to audio, identifies intent, transcribes meetings, and executes web automations — no wake word, no mandatory cloud, no visible traces.

• Continuous listening — mic always open, lightweight background processing
• Smart routing — classifies every audio chunk: command, meeting, or noise
• Meeting transcription — identifies speakers via DOM scraping (Meet/Teams) or OCR (Zoom)
• Keyword watching — acts silently when it hears trigger words during calls
• Web automations — controls the browser with human-like behavior via Ferrum
• Screen recording — captures, processes and exports feature demos
• Hidden terminal — Touch ID-protected control panel, invisible during screen share
• Discreet menu bar — neutral icon, no Dock, no notifications

Mic (always open)
    ↓
Cobra — VAD (is there a human voice?)
    ↓
Cheetah — real-time STT
    ↓
lightweight local filter (worth processing?)
    ↓
Rhino — Speech-to-Intent
    ↓
    ├── COMMAND      → ActionManager executes
    ├── MEETING      → buffer accumulates + OCR identifies speaker
    └── NOISE        → silently discarded

Swift is a thin native client. All business logic lives in Rails. Communication happens over a persistent WebSocket — faster than HTTP even on localhost.

Swift (thin client)                    Rails (orchestrator)
─────────────────────                  ────────────────────
AVAudioEngine                          ActionCable (WebSocket)
Cobra — VAD              ←——————————→  JardsChannel
Cheetah — STT                          ActionRegistry
Vision — OCR                           MeetingSession
ScreenCaptureKit                       BrowserPool + Ferrum
Menu bar + native UI                   picoLLM

Swift owns: audio capture, VAD, STT, OCR, native UI, WebSocket client Rails owns: all logic, state, browser automations, LLM decisions, persistence

IDLE → ALWAYS_ON → MEETING_MODE → SUMMARIZING → ALWAYS_ON

• IDLE — app paused, mic closed
• ALWAYS_ON — listening and routing commands
• MEETING_MODE — accumulating transcript, commands disabled
• SUMMARIZING — processing buffer with picoLLM

• AVAudioEngine — audio capture
• Picovoice (Cobra, Cheetah, Rhino, Orca, Falcon) — voice pipeline
• Vision Framework — on-device OCR to identify speakers
• ScreenCaptureKit — screen recording
• LocalAuthentication — Touch ID for hidden terminal
• NSStatusItem — menu bar
• URLSessionWebSocketTask — WebSocket client

• Action Cable — WebSocket server
• Ferrum — browser automation
• Sidekiq — job queue
• Browser::ActionRegistry — automation registry
• picoLLM — local LLM for decisions

The macOS orange microphone indicator is system-enforced for privacy — it cannot be removed. Nobody knows what the mic is doing though.

• macOS 13+
• Xcode Command Line Tools
• Ruby 3.2+
• Chrome installed

# 1. System dependencies
xcode-select --install
brew install xcode-build-server redis
brew services start redis

# 2. Rails
cd rails
bundle install
cp .env.example .env
# fill in PICOVOICE_KEY → console.picovoice.ai (free tier)
rails db:create db:migrate
rails s

# 3. Swift app
# open Jards.xcodeproj in Xcode or Zed
# cmd+R to run

xcode-build-server config -scheme Jards -project Jards.xcodeproj

// .zed/settings.json
{
  "languages": {
    "Swift": {
      "enable_language_server": true,
      "language_servers": ["sourcekit-lsp"],
      "format_on_save": "on"
    }
  }
}

"record screen"        → ScreenRecordingAction
"start meeting"        → MeetingTranscriptionAction
"summarize"            → SummarizeAction
"stop"                 → StopCurrentAction
"take screenshot"      → ScreenshotAction

linkedin.connect       → Browser::Actions::Linkedin::Connect
linkedin.send_message  → Browser::Actions::Linkedin::SendMessage
gmail.send             → Browser::Actions::Gmail::SendEmail
generic.screenshot     → Browser::Actions::Generic::Screenshot

• Swift app base + menu bar
• Cobra → Cheetah pipeline running
• WebSocket connection Swift ↔ Rails (Action Cable)
• ActionRegistry with first voice commands
• MeetingMode + OCR (Google Meet)
• MeetingMode + OCR (Zoom)
• HumanTyper + Ferrum (LinkedIn wrapper)
• Hidden terminal with Touch ID
• ScreenCaptureKit + post-processing pipeline
• picoLLM for ambiguous decisions
• Screen recording with automatic captions
• BrowserPool with anti-bot behavior

Jards doesn't exist. It's just a system utility.