Developer bmartin2000 has released AFK.exe, a cozy 3D simulation that rewards developers for physically leaving their workspace. Submitted for the Hacktoberfest 2026 Open-Source AI Challenge under the theme 'Touch Grass,' the project leverages local AI to verify outdoor activity without relying on cloud services. The core mechanic is simple: coding contracts drain energy, and only a GPS-qualified walk restores it, allowing players to upgrade their virtual office with plants and better equipment.

Local Inference Over Cloud Dependencies

The technical backbone of AFK.exe is a local implementation of Qwen3.5 4B, served via Ollama. This open-weight multimodal model performs two critical functions: validating nature photos submitted from the companion phone PWA and providing developer tools like an AI Debugger and Code Optimizer. By running inference on the user's own hardware, the project eliminates API costs and ensures that location data and personal photos never leave the device, a significant privacy win for developers wary of third-party tracking.

The Technical Stack and Workflow

Built with React, TypeScript, and Vite, the frontend renders the office using procedural Three.js geometry via React Three Fiber. Data persistence relies on IndexedDB for storing game progress, walk records, and photo evidence. The mobile component uses GPS to calculate walk distance, requiring a minimum of 10 minutes and 150 meters from the start point. When the user returns, the phone transfers a walk summary and re-encoded photos to the desktop via LAN or file transfer, where the local AI performs a plausibility check against mission criteria.

Deterministic Checks for Fair Play

While AI handles photo validation and code suggestions, the project maintains strict deterministic logic for core gameplay elements. Coding contracts are verified using authored test fixtures, ensuring that skill progression is based on actual code correctness rather than model hallucinations. Similarly, walk qualification is calculated using raw GPS data on the phone, independent of the language model. This hybrid approach prevents the 'black box' problem often seen in AI-gamified productivity tools, ensuring that rewards are earned through verifiable actions.

Key Takeaways

  • AFK.exe is MIT-licensed and runs entirely locally using Ollama and Qwen3.5 4B.
  • The project requires no account, cloud API keys, or internet connection after initial setup.
  • Outdoor verification requires a 10-minute walk of at least 150 meters, tracked via GPS.
  • AI-generated code suggestions are displayed as text only and are never executed automatically.

The Bottom Line

AFK.exe proves that local AI can solve real behavioral problems without sacrificing privacy or performance. It’s a clever infrastructure play that turns the often-neglected hardware capability of modern laptops into a personal wellness coach, forcing developers to step away from the screen to progress in their digital career.