One intelligent CEO agent that turns owner goals into auditable, evidence-backed work — delegated across a modular fleet of 270+ specialist agents, with a production-ready macOS voice assistant powered by Google Gemini Live API.
Live Dashboard → · Jarvis Voice Studio → · Architecture → · Quick Start →
CEO OS is a local-first, production-grade personal AI operating system designed for founders, executives, and power users who want a unified AI layer over their entire digital life.
It's not a chatbot. It's not a wrapper around OpenAI. It's an autonomous enterprise system with:
- 🧠 CEO ReAct Agent — multi-turn reasoning with XML thought scratchpads, tool dispatch, and reflection
- 🎙️ Jarvis — production-ready macOS voice assistant with local wake-word, zero idle costs, and Gemini Live bidirectional streaming
- 🖥️ CUA Desktop Automation — macOS Computer Use Agent that sees, clicks, and types
- 🌐 Hermes Multi-Agent Framework — orchestrates 270+ specialist agents from the Agency roster
- 🔌 gstack Integration — full software development workflow: Plan → Implement → Review → QA → Security → Release
- 📊 Cybernetic Dashboard — live chat interface with real-time task monitoring, voice control, and tool execution telemetry
"Jarvis, open YouTube" → Jarvis says "Opened sir, YouTube is now loading" → YouTube opens in Safari
Jarvis is a production macOS desktop voice assistant that runs 24/7 with zero idle API costs:
| Feature | Detail |
|---|---|
| Wake Word | Local openWakeWord ONNX inference — "Jarvis", "Computer", "Friday", "Hey Nova" |
| Connection Model | Gemini Live WebSocket opened only after wake word — $0 idle cost |
| Audio | Bidirectional PCM16 at 16kHz via sounddevice, native macOS Core Audio |
| Auth | Google Cloud Service Account JSON → Vertex AI OAuth2 (never exposed to frontend) |
| Voice | 5 Gemini prebuilt voices: Puck, Charon, Kore, Fenrir, Aoede |
| Barge-In | Instant cancellation — speaking interrupts playback via generation ID |
| macOS Tools | AppleScript automation: open URLs, launch apps, control Spotify, check system stats |
| Timeout | Configurable inactivity disconnect (15s–120s) |
| Configuration | All settings editable via the web dashboard |
┌─────────────────────────────────────────────────────────┐
│ JARVIS FLOW │
│ │
│ Mic → openWakeWord ──► "Jarvis" detected │
│ │ │
│ Connect Gemini Live WebSocket │
│ │ │
│ Bidirectional PCM Audio Streaming │
│ ┌─────────┴──────────┐ │
│ Speak Execute Tool │
│ │ │ │
│ SpeechSynthesis AppleScript/ │
│ (Jarvis replies) osascript │
│ │ │
│ Inactivity Timer → Disconnect │
└─────────────────────────────────────────────────────────┘
The CEO agent uses a multi-turn ReAct loop with:
- XML Thought Scratchpad —
<thought>,<tool_call>,<observation>,<reflection>,<final_answer> - Capability Router — matches natural language to the best specialist agent or tool
- ReAct Reflection — validates outputs before surfacing to the user
- Streaming Responses — typed progressively in the dashboard chat interface
# CEO ReAct Loop (simplified)
while not done:
thought = llm.think(context, tools)
if thought.has_tool_call:
output = tool_registry.execute(thought.tool_call)
context.add_observation(output)
elif thought.has_final_answer:
return thought.final_answerHermes manages a lazy-loaded roster of 270+ specialist Agency agents:
| Tool | Description |
|---|---|
agency_agents_search |
Semantic search across the full agent roster |
agency_agents_inspect |
Load and inspect a specific agent's skills |
agency_agents_load |
Activate selected agents for a task |
agency_agents_delegate |
Delegate work to the loaded agent fleet |
The CEO never loads all 270 agents at once. It searches, selects, and loads only what's needed — keeping context windows tight and costs minimal.
The Computer Use Agent (CUA) enables autonomous macOS control:
- Screenshot + Vision — captures and interprets the screen
- AppleScript/osascript — safe macOS automation with typed permissions
- App Control — launch, focus, type, click, scroll
- Browser Automation — Playwright-based web control
- Permission Matrix — every tool has ALLOW / ASK / DENY policy
CEO OS integrates gstack as a first-class development workflow provider:
CEO Agent
└── Development Director
└── gstack Skills Router
├── Planning
├── Implementation
├── Code Review
├── QA & Testing
├── Security Audit
└── Release Management
└── Coding Agent / Codex
└── Repository
The Cybernetic Dashboard at http://localhost:3000 provides:
- Executive Chat Deck — full chat-like interface with ReAct thought traces
- Voice Control — click mic or say "Jarvis" to send voice commands
- Live Task Feed — real-time WebSocket task cards with step-by-step execution
- Jarvis Voice Studio — configure wake word, model, voice, tools, and auth
- CUA Desktop Host — visual computer use sessions
- Agent Workforce — browse and inspect the 270+ specialist roster
- Memory Vault — persistent context and knowledge management
- System Health — infrastructure metrics and connection status
CEO-OS/
├── apps/
│ ├── api/ # FastAPI backend
│ └── dashboard/ # Next.js 14 frontend
│
├── core/ # CEO contracts, routing, orchestration
├── ceo_agent/ # ReAct reasoning engine
├── jarvis/ # Voice assistant (Gemini Live API)
├── hermes/ # Multi-agent orchestration framework
├── gstack/ # Software development workflow
│
├── agents/ # Domain agent implementations
├── tools/ # Capability implementations
├── integrations/ # External system connectors
│
├── memory/ # Persistent context and knowledge
├── computer/ # CUA desktop automation
├── browser/ # Playwright browser control
├── voice/ # TTS/STT providers
├── communications/ # Email, calendar, messaging
│
├── skills/ # Versioned reusable procedures
├── infrastructure/ # Docker, deployment configs
└── tests/ # 189 passing tests
- ✅ CEO kernel depends on contracts and registries, never provider SDKs directly
- ✅ External effects pass through typed capabilities, permissions, and audit events
- ✅ Model, vector, telephony, browser, storage providers are replaceable
- ✅ Untrusted external content is data, never instructions or authority
- ✅ Secrets referenced by credential identifiers — never in prompts or logs
- ✅ Default to workspace-scoped access and least privilege
- ✅ R2–R4 actions require policy evaluation; irreversible actions require approval
- Python 3.12+
uv(Python package manager)- Node.js 20+
- Docker + Docker Compose
- macOS (for Jarvis voice features)
- Google Cloud project with Vertex AI enabled (for Jarvis)
git clone https://github.com/abdullah917071/CEO-OS.git
cd CEO-OS
cp .env.example .env
# Edit .env with your API keysmake installmake infra-up# Option A: All services
make dev
# Option B: Individual services
make api # FastAPI backend at http://localhost:8000
make dashboard # Next.js dashboard at http://localhost:3000- Navigate to
http://localhost:3000/jarvis - Go to Google Cloud & Auth tab
- Paste your Google Cloud Service Account JSON
- Click Save & Store Securely
- Go to Overview and click Activate Gemini Live
- Say "Jarvis" — it will detect locally and connect
make check # Runs all 236 tests + full Ruff lint + strict Mypy + Next.js build + wheel buildAll configuration is available via the web dashboard at http://localhost:3000/jarvis or directly in .env:
| Variable | Description | Default |
|---|---|---|
OPENROUTER_API_KEY |
LLM provider for CEO agent | Required |
OPENROUTER_MODEL |
Model for reasoning | anthropic/claude-3.5-sonnet |
JARVIS_WAKE_WORD |
Wake word to activate Jarvis | jarvis |
JARVIS_INACTIVITY_TIMEOUT |
Seconds before auto-disconnect | 60 |
JARVIS_GEMINI_VOICE |
Prebuilt voice name | Puck |
GOOGLE_CLOUD_PROJECT |
GCP project ID for Vertex AI | From service account |
HERMES_AGENTS_PATH |
Path to Agency agent roster | ./agency |
- Service Account JSON stored with
0600filesystem permissions — never exposed to JavaScript - Private keys never placed in prompts, logs, or environment variables
- Tool permissions enforced at runtime: ALLOW / ASK / DENY per tool
- Audit trail — every tool execution logged with actor, arguments, and output
- Workspace-scoped access — default least privilege model
# Run full test suite
uv run pytest
# Run specific suites
uv run pytest tests/test_jarvis.py # Voice assistant tests
uv run pytest tests/test_ceo_agent.py # ReAct reasoning tests
uv run pytest tests/test_hermes_agent.py # Multi-agent tests
uv run pytest tests/test_gstack.py # Dev workflow tests
uv run pytest tests/test_import_smoke.py # Module import smoke tests
# Run complete quality gate checks
make check # lint + type check + tests + frontend build + wheel packagingCurrent Status: 236 tests passing ✅
- CEO ReAct agent with XML thought traces
- Basic tool dispatch and task management
- Dashboard with chat interface
- Local wake-word detection (zero idle cost)
- Gemini Live bidirectional WebSocket
- Safe macOS tool execution
- Web-based configuration dashboard
- 270+ specialist agent roster
- Lazy loading and semantic search
- Agent delegation and orchestration
- Development workflow provider
- Plan → Implement → Review → QA → Release pipeline
- Modular, replaceable architecture
- Proactive CEO intelligence
- Multi-model routing
- Advanced memory and context management
- Mobile companion app
| Layer | Technology |
|---|---|
| Backend | Python 3.12, FastAPI, asyncio |
| Package Management | uv |
| Frontend | Next.js 14, React 18, TypeScript |
| Styling | CSS Modules (no Tailwind dependencies) |
| LLM Providers | OpenRouter (CEO), Google Vertex AI (Jarvis) |
| Voice | Google Gemini Live API, openWakeWord ONNX, sounddevice |
| Database | SQLite (local), PostgreSQL (production) |
| Cache | Redis |
| Search | Qdrant (vector) |
| Infrastructure | Docker Compose |
| Testing | pytest, Playwright |
| Linting | Ruff, mypy, ESLint |
apps/ Deployable API and dashboard
core/ CEO contracts, routing, orchestration
ceo_agent/ ReAct reasoning engine with XML scratchpad
jarvis/ Production macOS voice assistant (Gemini Live)
hermes/ Multi-agent orchestration framework
gstack/ Software development workflow integration
agents/ Domain agent implementations
tools/ Typed capability implementations
integrations/ External system connectors (Google, Meta, etc.)
memory/ Persistent context and knowledge management
computer/ CUA desktop automation (macOS)
browser/ Playwright browser automation
voice/ TTS/STT provider abstractions
communications/ Email, calendar, Slack, messaging
skills/ Versioned reusable procedures
infrastructure/ Docker, nginx, deployment configs
plans/ Architecture decision records
docs/ Numbered architecture source of truth
tests/ Unit, integration, contract, and evaluation suites
See AGENTS.md for the engineering rules and architecture guardrails. Read CURRENT_STATE.md for verified capabilities and PLANS.md for the roadmap.
Before contributing:
- Read
AGENTS.md,CURRENT_STATE.md,PLANS.md, and relevantdocs/documents - Inspect the current implementation
- Work on one roadmap milestone at a time
- Add/update tests, run
make check, fix failures - Update
CURRENT_STATE.mdonly with behavior that was verified
MIT License — see LICENSE for details.
Built with ❤️ for the AI-native founder
CEO OS · Autonomous Enterprise AI · Powered by Gemini Live + OpenRouter
CEO OS · Autonomous Enterprise AI · Powered by Gemini Live + OpenRouter