zero-employee-orchestrator 0.1.4

Zero-Employee Orchestrator – AI Orchestration Platform for multi-agent business automation with human approval and full auditability

Language: English | 日本語 | 简体中文 | 繁體中文 | 한국어 | Português (Brasil) | Türkçe

Zero-Employee Orchestrator

The AI Meta-Orchestrator — Design · Execute · Verify · Improve

---

The orchestrator of orchestrators. One platform to command every AI framework, every LLM, every tool — under human approval and full audit.

ZEO doesn't replace your AI tools — it unifies them. Connect CrewAI, AutoGen, LangChain, Dify, Claude Cowork, n8n, Zapier, and 34+ business apps under a single approval gate, audit trail, and security layer. Define workflows in natural language. Let multiple AI agents plan, execute, verify, and improve — all under your control.

ZEO itself is free and open source. LLM API costs are paid directly by users to each provider.

---

Getting Started

Choose your path:

Method	Best for	Time	API key needed?
Desktop App	Non-technical users	2 min	No (subscription mode)
CLI (pip install)	Developers	2 min	No (subscription or Ollama)
Docker	Self-hosting / production	5 min	No (subscription or Ollama)

System Requirements: Python 3.12+ (CLI), Node.js 22+ (frontend dev), 4 GB RAM minimum. Ollama local models need 8 GB+ RAM.

---

🖥️ Download Desktop App

Pre-built desktop installers are available on the Releases page.

OS	File	Description
Windows	-setup.exe	Windows installer (x64)
macOS	.dmg	macOS Universal (Intel + Apple Silicon)
Linux	.AppImage	Portable (no install needed, amd64)
Linux	.deb / .rpm	Debian/Ubuntu / Fedora/RHEL (amd64/x86_64)

After installation, a setup wizard will guide you through:

1. Language — Choose English, 日本語, 中文, 한국어, Português, or Türkçe (changeable later in Settings)
2. LLM provider — Pick how the AI runs (no API key needed for subscription mode)
3. First task — Start using the platform immediately

---

🚀 Quick Start (CLI)

Step 1: Install

# PyPI (recommended)
pip install zero-employee-orchestrator

# or from source
git clone https://github.com/OrosiTororo/Zero-Employee-Orchestrator.git
cd Zero-Employee-Orchestrator && pip install .

# or Docker (see Docker section below for details)
docker compose -f docker/docker-compose.yml up -d

Step 2: Configure

Pick one of these options:

# Option A: No API key needed — uses free web AI services via g4f
zero-employee config set DEFAULT_EXECUTION_MODE subscription

# Option B: Fully offline — local models via Ollama (no internet needed)
zero-employee config set DEFAULT_EXECUTION_MODE free
zero-employee pull qwen3:8b

# Option C: API key — best quality, pay-per-use to provider
zero-employee config set OPENROUTER_API_KEY <your-key>  # or GEMINI_API_KEY, etc.

ZEO itself is free. LLM costs (if any) are paid directly to each provider. See USER_SETUP.md for all options.

Step 3: Start

# Option A: start script (starts both backend + frontend automatically)
./start.sh                       # macOS / Linux
.\start.ps1                      # Windows (PowerShell)
# → Open http://localhost:5173

# Option B: Manual start
zero-employee serve              # Start the API server (port 18234)
cd apps/desktop/ui && pnpm dev   # Start the frontend (port 5173) in another terminal
# → Open http://localhost:5173

# Option C: Chat mode only (no Web UI needed)
zero-employee chat               # Default settings
zero-employee local --model qwen3:8b  # Ollama

Note: zero-employee serve starts the API server only. The Web UI runs separately on port 5173. Use start.sh (or start.ps1 on Windows) for the easiest setup.

Step 4: Verify

zero-employee health              # Check server status
zero-employee models              # List available models
zero-employee config list         # Review your settings

Changing Language

The default language is English. Change it system-wide (CLI, AI responses, and Web UI all switch together):

# At startup
zero-employee chat --lang ja      # Japanese
zero-employee chat --lang zh      # Chinese
zero-employee chat --lang ko      # Korean
zero-employee chat --lang pt      # Portuguese
zero-employee chat --lang tr      # Turkish

# Persistently (saved to ~/.zero-employee/config.json)
zero-employee config set LANGUAGE ja

# At runtime (inside chat mode)
/lang en                          # Switch to English
/lang ja                          # Switch to Japanese
/lang zh                          # Switch to Chinese
/lang ko                          # Switch to Korean
/lang pt                          # Switch to Portuguese
/lang tr                          # Switch to Turkish

In the desktop app, change language anytime via Settings.

---

🐳 Docker

API + Frontend (recommended)

docker compose -f docker/docker-compose.yml up -d
# → Open http://localhost:5173

This starts three services: API server (port 18234), Frontend (port 5173), and a background worker.

Note: Requires SECRET_KEY environment variable. Generate one: python -c "import secrets; print(secrets.token_urlsafe(32))"

API only

docker compose up -d
# → API available at http://localhost:18234/api/v1/

This starts only the API server. Use this with the Desktop App or your own frontend.

---

The Guides

Quickstart Guide First workflow, CLI basics.	Architecture Deep Dive 9-layer architecture, DAG, Judge Layer.	Security Guide Prompt defense, approval gates, sandbox.

---

📦 What's Inside

Zero-Employee-Orchestrator/
├── apps/
│   ├── api/                  # FastAPI backend
│   │   └── app/
│   │       ├── core/               # Config, DB, security, i18n
│   │       ├── api/routes/         # 46 REST API route modules
│   │       ├── api/ws/             # WebSocket
│   │       ├── models/             # SQLAlchemy ORM
│   │       ├── schemas/            # Pydantic DTO
│   │       ├── services/           # Business logic
│   │       ├── repositories/       # DB I/O abstraction
│   │       ├── orchestration/      # DAG, Judge, state machine
│   │       ├── providers/          # LLM gateway, Ollama, RAG
│   │       ├── security/           # IAM, secrets, sanitize, prompt defense
│   │       ├── policies/           # Approval gates, autonomy boundaries
│   │       ├── integrations/       # Sentry, MCP, external skills, Browser Assist
│   │       └── tools/              # External tool connectors
│   ├── desktop/              # Tauri v2 + React UI
│   ├── edge/                 # Cloudflare Workers
│   └── worker/               # Background workers
├── skills/                   # 8 built-in skills
├── plugins/                  # 16 plugin manifests (10 general + 6 role-based)
├── extensions/               # 11 extension manifests
│   └── browser-assist/
│       └── chrome-extension/ # Chrome extension for Browser Assist
├── packages/                 # Shared NPM packages
├── docs/                     # Multi-language docs & guides
│   ├── ja-JP/                # 日本語
│   ├── zh-CN/                # 简体中文
│   ├── zh-TW/                # 繁體中文
│   ├── ko-KR/                # 한국어
│   ├── pt-BR/                # Português (Brasil)
│   ├── tr/                   # Türkçe
│   └── guides/               # Architecture, security, quickstart guides
└── assets/
    └── images/
        ├── guides/           # Guide header images
        └── logo/             # Logo assets

---

🏗️ 9-Layer Architecture

┌─────────────────────────────────────────┐
│  1. User Layer       — Natural language input          │
│  2. Design Interview — Requirements exploration        │
│  3. Task Orchestrator — DAG decomposition & scheduling  │
│  4. Skill Layer      — Specialized Skills + Context     │
│  5. Judge Layer      — Two-stage + Cross-Model QA       │
│  6. Re-Propose       — Rejection → dynamic DAG rebuild  │
│  7. State & Memory   — Experience Memory               │
│  8. Provider         — LLM Gateway (LiteLLM)           │
│  9. Skill Registry   — Publish / Search / Import        │
└─────────────────────────────────────────┘

---

🎯 Key Features

Core Orchestration

Feature	Description
Design Interview	Natural language requirements exploration and refinement
Spec / Plan / Tasks	Structured intermediate artifacts — reusable, auditable, reversible
Task Orchestrator	DAG-based planning with cost estimation and quality mode switching
Judge Layer	Rule-based first pass + Cross-Model high-accuracy verification
Self-Healing / Re-Propose	Automatic re-planning on failure with dynamic DAG reconstruction
Experience Memory	Learns from past executions to improve future performance

Meta-Orchestration (Orchestrator of Orchestrators)

Feature	Description
Agent Adapter	Integrate CrewAI, AutoGen, LangChain, Dify agents as ZEO sub-workers — all under approval gates and audit
iPaaS Bridge	Connect n8n / Zapier / Make workflows — ZEO becomes the judgment layer for existing automations
MCP Server	Model Context Protocol — any MCP-compatible tool becomes a ZEO tool automatically
App Connector Hub	34+ apps (Obsidian, Notion, Google Workspace, Microsoft 365, Slack, etc.) — connectors for connectors
Tool-of-Tools	Tools that connect to other tools (e.g., n8n connects to 400+ apps → ZEO connects to n8n → ZEO accesses all 400+ apps through one integration)
Multi-Model Freedom	26 model families via LiteLLM, Ollama local models, g4f web sessions — not locked to any single provider

Extensibility

Feature	Description
Skill / Plugin / Extension	3-tier extensibility with full CRUD management
Role-Based Plugin Packs	Pre-configured plugins per business role: Sales, Finance, HR, Legal, Marketing, Support
Natural Language Skill Generation	Describe in plain language → AI auto-generates (with safety checks)
Skill Marketplace	Community skill publishing, search, review, and installation
External Skill Import	Import skills from GitHub repositories
Self-Improvement	AI analyzes and improves its own skills (with approval)
Meta-Skills	AI learns how to learn (Feeling / Seeing / Dreaming / Making / Learning)

AI Capabilities

Feature	Description
Browser Assist	Chrome extension overlay — AI sees your screen in real-time
Media Generation	Image, video, audio, music, 3D — with dynamic provider registration
App Connector Hub	34+ apps (Obsidian, Notion, Google Workspace, Microsoft 365, etc.)
AI Tool Integration	55+ external tools across 21 categories
A2A Communication	Peer-to-peer agent messaging, channels, and negotiation
Avatar AI	Learns your decision patterns and evolves with you
Secretary AI	Brain dump → structured tasks, bridges you and the AI org
Operator Profile	Cowork-style about-me + global instructions — AI personalizes responses based on your role, priorities, and work style
Task Dispatch	Cowork Dispatch-inspired background tasks — fire-and-forget with status polling
Repurpose Engine	Auto-convert 1 content to 10 media formats

Security

Feature	Description
Prompt Injection Defense	5 categories, 28+ detection patterns
Approval Gates	14 categories — including tiered browser automation (10 levels: navigate < click < type < submit < login < payment)
File Sandbox	AI can only access user-permitted folders (default: STRICT)
Data Protection	Upload/download policy control (default: LOCKDOWN)
PII Protection	Auto-detect and mask 13 categories of personal information
IAM	Human/AI account separation, AI denied admin/secret access
Red-team Security	8-category, 20+ test self-vulnerability assessment

Operations

Feature	Description
Multi-model Support	Dynamic catalog, auto-fallback, per-task provider override
i18n	6 languages (EN / JA / ZH / KO / PT / TR) — UI, AI responses, CLI
Autonomous Operation	Docker / Cloudflare Workers — runs when your PC is off
24/365 Scheduler	9 trigger types: cron, ticket creation, budget threshold, etc.
iPaaS Integration	n8n / Zapier / Make webhook integration
Cloud Native	AWS / GCP / Azure / Cloudflare abstraction layer
Governance & Compliance	GDPR / HIPAA / SOC2 / ISO27001 / CCPA / APPI

---

🔒 Security

ZEO is designed security-first with multi-layered defense:

Layer	Description
Prompt Injection Defense	Detects and blocks instruction injection from external inputs (5 categories, 28+ patterns)
Approval Gates	14 categories of dangerous operations (send, delete, billing, permission changes) require human approval
Autonomy Boundaries	Explicitly limits what AI can do autonomously
IAM & Tool Permissions	Separate human/AI accounts; role-based tool permissions (5 default policies: secretary, researcher, reviewer, executor, admin) enforce least privilege per agent
Kill Switch	Emergency halt of all active executions via UI button or API (/kill-switch/activate). Blocks new executions until resumed.
Tiered Judge	Three-tier verification: LIGHTWEIGHT (rules only) → STANDARD (+policy) → HEAVY (+cross-model). Reduces cost for low-risk ops while maintaining full verification for high-risk ones.
Memory Trust	Experience Memory entries track source type, trust level (0.0-1.0), verification status, and expiry. Only trustworthy memories (≥0.7, not expired) are used.
Secret Management	Fernet encryption, auto-masking, rotation support
Sanitization	Auto-removal of API keys, tokens, and PII
Security Headers	CSP, HSTS, X-Frame-Options on all responses
Rate Limiting	slowapi-based API rate limiting
Audit Logging	All critical operations recorded (built-in from design, not bolted on)

See SECURITY.md for vulnerability reporting.

---

🖥️ CLI Reference

zero-employee serve              # Start API server
zero-employee serve --port 8000  # Custom port
zero-employee serve --reload     # Hot reload

zero-employee chat               # Chat mode (all providers)
zero-employee chat --mode free   # Free mode (Ollama / g4f)
zero-employee chat --lang en     # Language selection

zero-employee local              # Local chat (Ollama)
zero-employee local --model qwen3:8b --lang ja

zero-employee models             # List installed models
zero-employee pull qwen3:8b      # Download model

zero-employee config list        # Show all settings
zero-employee config set <KEY>   # Set a value
zero-employee config get <KEY>   # Get a value

zero-employee db upgrade         # Run DB migrations
zero-employee health             # Health check
zero-employee security status    # Security status
zero-employee update             # Update to latest version

---

🤖 Supported LLM Models

Managed via model_catalog.json — swap models without code changes.

Mode	Description	Examples
Quality	Highest quality	Claude Opus, GPT, Gemini Pro
Speed	Fast response	Claude Haiku, GPT Mini, Gemini Flash
Cost	Low cost	Haiku, Mini, Flash Lite, DeepSeek
Free	Free	Gemini free tier, Ollama local
Subscription	No API key needed	via g4f

Per-task provider override is supported — specify provider, model, and execution mode per task.

---

🧩 Skill / Plugin / Extension

3-Tier Extensibility

Type	Description	Examples
Skill	Single-purpose specialized processing	spec-writer, review-assistant, browser-assist
Plugin	Bundles multiple Skills	ai-secretary, ai-self-improvement, youtube
Extension	System integration & infrastructure	mcp, oauth, notifications, browser-assist

Generate Skills with Natural Language

POST /api/v1/registry/skills/generate
{
  "description": "A skill that summarizes long documents into 3 key points"
}

18 dangerous patterns are auto-detected. Only skills passing safety checks are registered.

---

🌐 Browser Assist

Chrome extension overlay chat — AI sees your screen in real-time and guides you.

• Overlay Chat: Chat UI directly on any website
• Real-time Screen Sharing: AI sees what you see (no manual screenshots)
• Error Diagnosis: AI reads error messages on screen and suggests fixes
• Form Assistance: Step-by-step field-by-field guidance
• Privacy-first: Screenshots processed temporarily, PII auto-masked, password fields blurred

Setup

1. Load extensions/browser-assist/chrome-extension/ in Chrome
   → chrome://extensions → Developer mode → "Load unpacked"
2. Click the chat icon on any website
3. Ask questions or share your screen with the screenshot button

---

🛠️ Tech Stack

Backend

• Python 3.12+ / FastAPI / uvicorn
• SQLAlchemy 2.x (async) + Alembic
• SQLite (dev) / PostgreSQL (production)
• LiteLLM Router SDK
• bcrypt / Fernet encryption
• slowapi rate limiting

Frontend

• React 19 + TypeScript + Vite
• shadcn/ui + Tailwind CSS
• TanStack Query + Zustand

Desktop

• Tauri v2 (Rust) + Python sidecar

Deploy

• Docker + docker-compose
• Cloudflare Workers (serverless)

---

❓ FAQ

Do I need API keys to start?

No. You can use subscription mode (no key needed) or Ollama for fully offline local AI. See the Quick Start section above.

How much does it cost?

ZEO itself is free. LLM API costs are paid directly by you to each provider (OpenAI, Anthropic, Google, etc.). You can also run completely free with Ollama local models.

Can I use multiple LLM providers simultaneously?

Yes. ZEO supports per-task provider override — you can use Claude for high-quality spec reviews and GPT for fast task execution in the same workflow.

Is my data safe?

ZEO is self-hosted by design. Your data stays on your infrastructure. File sandbox defaults to STRICT, data transfer defaults to LOCKDOWN, and PII auto-detection is enabled by default.

How is this different from AutoGen / CrewAI / LangGraph?

ZEO is a business workflow platform, not a developer framework. It provides human approval gates, audit logging, a 3-tier extensibility system, browser assist, media generation, and a complete REST API — all designed for running AI as an organization, not just chaining prompts.

---

🧪 Development

# Setup
git clone https://github.com/OrosiTororo/Zero-Employee-Orchestrator.git
cd Zero-Employee-Orchestrator
pip install -e ".[dev]"

# Start (hot reload)
zero-employee serve --reload

# Test
pytest apps/api/app/tests/

# Lint
ruff check apps/api/app/
ruff format apps/api/app/

---

🤝 Contributing

Contributions are welcome.

1. Fork → Branch → PR (standard flow)
2. Security issues: follow SECURITY.md for private reporting
3. Coding standards: ruff format, type hints required, async def

---

💜 Sponsors

This project is free and open source. Sponsors help keep it maintained and growing.

Become a Sponsor

---

🌟 Star History

---

📄 License

MIT — Use freely, modify as needed, contribute back if you can.

---

Zero-Employee Orchestrator — Run AI as an organization.
Built with security, auditability, and human oversight in mind.