Download
Google Gemini 3.7 Flash & Google ADK

The Autonomous Desktop AI Coworker that Sees, Thinks & Acts

Cup Work controls native Windows applications, automates document creation in MS Word and Excel, delivers grounded daily news briefings, displays on-screen scratchpads, and draws live whiteboard lectures directly over your desktop.

Terminal Setup
Windows 10 / 11 (64-bit)
FastAPI Python 3.12 Backend
Electron 34 Transparent Client
Cup Work Live Desktop Demo
Pillars of Intelligence

Engineered for Complete Desktop Autonomy

Cup Work bridges cutting-edge Gemini reasoning directly to the native Windows desktop API for seamless document authoring, real-time news retrieval, and workspace assistance.

MS Word & Office Document Automation

Autonomously clicks ribbon buttons, navigates menus, formats headings, inserts tables, and writes complete structured documents directly in Microsoft Word and Excel.

  • Auto-typing text, bullet points, and formatting
  • Automated mouse navigation & ribbon clicking

On-Screen Windows Scratchpad

A dedicated floating on-screen card that summarizes complex analyses, architectural insights, and runnable shell commands with 1-click clipboard copying.

  • Formatted Markdown cards with headers & code blocks
  • Runnable shell command deck with instant copy

Grounded News Reader & Daily Briefing

Searches breaking world affairs, tech headlines, and financial markets using Google Search grounding, delivering concise audio briefings and written reports.

  • Real-time Google Search news grounding
  • Voice narrated morning summaries & key takeaways

On-Screen Whiteboard & Drawing Tutor

Draws dynamic architectural diagrams, algorithms, and math explanations on a transparent overlay layered directly across your active desktop apps.

  • Auto-gliding camera with pan & zoom
  • Mermaid JS & custom SVG path renderers

Multimodal Screen Vision HUD

Understands visual elements in real time, projecting bounding boxes and pointer arrows directly onto UI buttons, menus, and code editors.

  • Sub-100ms frame coordinate localization
  • Win32 Accessibility tree integration

Multi-App Tasks & Todo Manager

Organizes cross-application workflows, schedules daily milestones, prioritizes tasks, and manages your permanent SQLite activity log.

  • Dynamic priority tagging & completion tracking
  • Temporal memory persistence across restarts
Currently Running on GCP (Google Cloud)
Google Cloud Run
Google ADK Multi-Agent Core

Interactive Multi-Agent Execution Flow

Explore how Cup Work orchestrates Gemini 3.7 Flash sub-agents, memory recall, verification guardrails, and real-time WebSocket delivery to your desktop. Click any node to inspect its specification.

Workflow:
Gemini 3.7 Flash Reasoning Core Gemini 3.1 TTS Cup Work User Voice 1 MemoryManager SQLite Storage AdkRunner Session Service Root Agent Router ADK Coordinator TOOLS [ALL] Chrome Agent TOOLS WinExecutor Word/Excel UIA TOOLS Screen Annotator HUD TOOLS OnScreen Whiteboard SVG TOOLS Clarification Notes Agent NOTE TOOLS Research News Agent SEARCH Scratchpad Card Deck PAD TOOLS Verification Loop GoalVerifier Human-In-The- Loop Dialogs Gemini TTS Streamer 5 EventBus Commentary & Events Electron Bridge WebSocket :8765
Root Orchestrator (Router) Gemini 3.7 Flash

Top-level ADK agent that classifies user intent, queries the temporal SQLite memory manager for context, and dynamically delegates sub-tasks to specialized domain agents.

Model Runtime
gemini-3.7-flash (Vertex AI)
Assigned Tools
Agent Delegation, Memory Recall
Input Protocol
User Prompt + Temporal Memory
Output Protocol
Sub-Agent Dispatch & TTS Stream
Release Hub

Download Cup Work for Windows

Official pre-compiled production installer and developer bundles.

Production Release

Cup Work Desktop Installer

Includes the Electron 34 desktop client, transparent screen overlay system, Win32 Automation drivers, and integrated Python bridge.

GitHub Repository
Filename Cup-Work-Setup-1.0.0.exe
Version v1.0.0
File Size 104.4 MB
Target OS Windows 10 / 11 (64-bit)
License MIT Open Source
Signature Status Local Build Package
Developer Setup

2-Minute Quick Start Guide

Spin up the Python Brain server and Electron client locally in two terminal sessions.

Terminal 1 (Python Brain Server)
$cd backend
$python -m venv .venv
# Activate virtual environment
$.\.venv\Scripts\activate
$pip install -r requirements.txt
# Start FastAPI Brain on port 8765
$python main.py
Terminal 2 (Electron Desktop Client)
$npm install
# Compile TypeScript main & preload
$npm run build:electron
# Launch Desktop Companion & Transparent Screen Canvas
$npm start
Judge Q&A & Technical Deep-Dive

Key Architectural Decisions & Real Answers

Direct technical insights on Google Gemini 3.7 Flash integration, ADK multi-agent routing, safety guardrails, low-latency streaming voice, and desktop automation benchmarks.

Why Gemini 3.7 Flash & Google ADK over monolithic general LLMs?

Gemini 3.7 Flash provides sub-second multimodal vision reasoning and sub-pixel spatial accuracy critical for desktop UI localization. Paired with the Google Agent Development Kit (ADK), it enables modular specialization: rather than overburdening a single prompt with 40+ desktop tools, the Root Agent dynamically delegates to domain-isolated agents (WinExecutor, Scratchpad, Whiteboard, Research), minimizing context degradation and eliminating tool hallucinations.

How does Cup Work reliably automate MS Word & Excel without fragile coordinate clicking?

We implement a hybrid Automation Engine: native Win32 UI Automation (UIA) accessibility trees coupled with Gemini 3.7 Vision OCR. If an element lacks accessibility tags, vision localization calculates exact bounding boxes. Crucially, every action is visually validated by GoalVerifier before proceeding to the next step, allowing automatic mid-flight self-correction and retries.

How is real-time voice streaming with emotional inflection achieved?

The voice pipeline integrates gemini-3.1-flash-tts-preview over Server-Sent Events (SSE). Text tokens stream in parallel with task execution and synthesize into 24kHz PCM audio buffers. Emotion tags like [excitedly], [thoughtful], and [firm] are parsed on the fly, delivering expressive companion speech with sub-500ms time-to-first-audio.

What safety guardrails and Human-in-the-Loop (HITL) protections prevent destructive actions?

High-impact operations (such as file deletions, executing raw terminal commands, outbound email dispatch, or payment portals) automatically trigger the ask_human_tool HITL Dialog. The Electron client renders a blocking modal card presenting the exact command, arguments, and affected paths, requiring explicit user confirmation before the WinExecutor proceeds.

How is cross-session memory managed without compromising enterprise privacy?

Cup Work utilizes a local-first temporal memory engine powered by embedded SQLite. User preferences, workspace configurations, and conversational summaries are auto-summarized and indexed entirely on the local machine. Zero document content or confidential telemetry is stored on external servers.

How is the dual-engine architecture packaged and deployed for production?

The Python backend is packaged via Docker and deployable to Google Cloud Run or run locally as a FastAPI runtime. The Electron 34 desktop client uses transparent hardware-accelerated canvas overlays (overlay.html) and communicates with the Python brain over a local WebSocket on port 8765 with sub-millisecond IPC latency.