Skip to main content

Zero-Token Desktop Operator AI (ZT-Operator)

On-Device, Privacy-First Agentic Desktop Co-Pilot Powered by Gemma 4

Executive Overview

The Zero-Token Desktop Operator (ZT-Operator) is a local-first automation assistant that executes everyday desktop workflows directly on consumer workstation hardware. By combining Google’s Gemma 4 (26B A4B / 31B) with native local tool integration, ZT-Operator delivers zero-latency, zero-cloud-cost, and 100% data-sovereign productivity automation.

Designed with a high-contrast, minimalist Black & White Workstation UI, ZT-Operator gives users instant clarity, one-click workflow triggers, live hardware telemetry, and an uncompromising security model.

Key Capabilities

  • 100% Data Sovereignty: Personal emails, calendar events, local documents, and system metrics never leave your device.
  • Minimalist Black & White UI: Ultra-readable, high-contrast monochrome design inspired by professional developer workstations.
  • 8 Native Workstation Tools: Live psutil system metrics monitoring, top RAM/CPU process inspector, local PDF/TXT parsing, local ICS calendar parsing, directory audits, daily morning briefings, sandboxed file organization, and terminal execution.
  • Human-In-The-Loop (HITL) Security: Dangerous or state-changing actions (file mutations, shell scripts) trigger an interactive authorization modal before execution.
  • $0.00 Token Cost: Runs locally via Ollama / LiteRT / vLLM.

Architecture & Data Flow

┌────────────────────────────────────────────────────────────────────────┐
│                        LOCAL USER WORKSTATION                          │
│                                                                        │
│   ┌────────────────────────────────────────────────────────────────┐   │
│   │   Minimalist B&W Workstation UI (HTML5 / Tailwind CSS)         │   │
│   │   - Real-time CPU, RAM & Disk hardware meters                  │   │
│   │   - Live top memory/CPU running process feed                   │   │
│   │   - Tools Playground & One-Click Workflow triggers             │   │
│   │   - Human-In-The-Loop (HITL) Interactive Confirmation Modal    │   │
│   └──────────────────────────────┬─────────────────────────────────┘   │
│                                  │ HTTP / JSON API (localhost:8000)    │
│                                  ▼                                     │
│   ┌────────────────────────────────────────────────────────────────┐   │
│   │   FastAPI Agent Engine (agent_server.py)                       │   │
│   │   - 8 System Tools & Safe Sandboxed Execution Matrix           │   │
│   │   - Deterministic HITL Gate & Authorization History Log        │   │
│   └──────────────────┬───────────────────────────┬─────────────────┘   │
│                      │ Function Calling          │ Local System Read/Write │
│                      ▼                           ▼                     │
│   ┌──────────────────────────────┐   ┌─────────────────────────────┐   │
│   │ Gemma 4 (Local Ollama Engine)│   │ Local System Resources      │   │
│   │ - Gemma-4-26B-A4B-IT         │   │ - System Hardware (psutil)  │   │
│   │ - Native JSON Tool Calling   │   │ - Local Filesystem (~/)     │   │
│   │ - Context Window: Up to 256K │   │ - iCal Calendars (.ics)     │   │
│   └──────────────────────────────┘   └─────────────────────────────┘   │
└────────────────────────────────────────────────────────────────────────┘

Tool Safety Protocol (Human-in-the-Loop Matrix)

Tool Name Action Type Confirmation Required? Action Taken
get_system_metrics Read-Only No (Auto-Executed) Live CPU %, Memory %, Disk %, process count
get_top_processes Read-Only No (Auto-Executed) Lists top RAM/CPU consumer apps
get_daily_briefing Read-Only No (Auto-Executed) Compiles schedule, health & running apps
parse_local_document Read-Only No (Auto-Executed) Extracts text from local PDFs/TXT files
get_calendar_events Read-Only No (Auto-Executed) Parses local .ics calendar files
list_directory Read-Only No (Auto-Executed) Audits folder contents & file sizes
organize_files Mutation YES (HITL Prompt) Moves/renames files on disk
execute_shell_script Mutation YES (HITL Prompt) Runs sandboxed terminal command

Benchmarks & Performance Profile (Tested on M4 MacBook Pro / RTX 4090)

Metric Gemma 4 26B A4B (Q4_K_M) Cloud API Alternative
API Cost per 1M Tokens $0.00 $3.00 - $15.00
Time to First Token (TTFT) 120 ms 450 - 900 ms
Data Transmitted Off-Device 0 Bytes Megabytes of personal data
Offline Operation 100% Functional Fails completely

Quickstart & Installation

1. Prerequisites

  • Python 3.9+ installed
  • Ollama installed (for Gemma 4 local inference)

2. Install Python Dependencies

pip install -r requirements.txt

3. Launch Local Model with Ollama

ollama run gemma4:26b

4. Start ZT-Operator Backend

python agent_server.py

(Or double-click start_workstation.bat on Windows).

5. Open Workstation UI

Navigate to http://localhost:8000 in your web browser.


Project Structure

zt-operator/
├── README.md               # Hackathon documentation, architecture, benchmarks
├── agent_server.py         # FastAPI Agent Engine, 8 System Tools & HITL Manager
├── index.html              # High-contrast Black & White Workstation UI
├── requirements.txt        # Backend dependencies (fastapi, psutil, uvicorn, etc.)
├── sample_calendar.ics     # Sample iCalendar file with test events
├── sample_document.txt     # Sample document file for parsing
├── start_workstation.bat   # Windows one-click launcher
└── .gitignore              # Clean repository rules

License

MIT License - Open Source for Hackathon Evaluation.

Built with

pythonjavascriptgeminiollamagemma

Prize Categories Entered

  • Deepmind Google

    Best Use of Gemma 4

    Submitted