🌿 Side Quest: An Open-Source AI Agent That Turns Screen Time into Real-World Adventures

🌿 Side Quest: An Open-Source AI Agent That Gets You Off Your Screen

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.

What I Built

We spend hours staring at screens, often forgetting to step outside, take a walk, or notice the world around us.

Side Quest is an open-source AI-powered outdoor companion that turns excessive screen time into small real-world adventures.

Instead of simply telling you to take a break, Side Quest creates a personalized 20-minute quest based on your local weather, sunset time, and nearby green spaces.

🌱 What it does:

• Detects screen time: Tracks activity inside the PWA, with an optional lightweight desktop watcher for system-wide activity.
• Generates outdoor quests: Uses open-weight AI models to suggest simple activities, like finding three different red leaves or exploring a nearby park.
• Adapts to conditions: Checks weather, temperature, rain, storms, and sunset before suggesting a quest.
• Gets you away from the screen: Switches to a dark, voice-friendly interface so you can complete your quest without constantly looking at your phone.
• Verifies your adventure: Lets you submit a proof photo for AI-assisted verification and earn a stamp.
• Keeps your memories local: Stores your quest history, photos, and statistics in your browser.

When outdoor conditions aren't suitable, Side Quest suggests a safer indoor-adjacent alternative, such as a balcony or covered area.

The goal is simple: less scrolling, more exploring.

Demo

🎥 Watch the Demo: Side Quest — Video Demo

🌐 GitHub Repository: Tusharkhadde/side-quest

💻 GitHub Repository: Tusharkhadde/side-quest

The project is open source under the MIT License. You can run it locally, experiment with different open-weight models, or self-host the inference layer.

How I Built It

Side Quest is built with Next.js, TypeScript, browser APIs, and an AI agent pipeline designed to work with multiple model providers.

The open-source AI stack

• Gemma 4 26B A4B: The default model for generating quests and checking proof photos through the Gemini API.
• Gemma 4 31B: A text-generation fallback through DigitalOcean's serverless inference.
• NVIDIA Nemotron Nano 12B v2 VL: An open-weight vision-language model for photo verification through DigitalOcean.
• Ollama: An optional route for running compatible models locally.
• TabPFN: An optional experiment that predicts when a user is most likely to accept a quest based on their exported journal data.

How the agent works

1. Gather local context, including approximate location, weather, sunset time, and nearby parks or paths.
2. Apply code-based safety rules to determine which types of quests are appropriate.
3. Ask the selected open-weight model to generate a structured quest with clear steps and a safety note.
4. Validate the response and retry or switch providers if generation fails.
5. Let the user complete the quest, submit a photo if they choose, and earn a stamp.

The application uses Open-Meteo for weather information and OpenStreetMap data for nearby places.

It also includes a lightweight desktop watcher that detects extended periods of computer activity without sending activity data to a remote server.

Why Does Open Innovation Matter?

For a project focused on personal habits and privacy, having control over the AI stack matters.

Open-weight models make it possible to experiment with different inference providers, change models without rebuilding the application, and eventually run inference entirely on local hardware.

With Side Quest:

• No single-provider lock-in: The provider chain can switch between hosted inference, compatible endpoints, Ollama, and a deterministic demo fallback.
• More control over privacy: Location processing happens on the device, and self-hosted inference can keep model inputs on your own hardware.
• Transparent experimentation: Developers can inspect the prompts, validation schemas, provider integrations, and safety logic.
• Room to improve: Quest completion history can support future experiments with personalization and fine-tuning.

One important distinction: open weights don't automatically make hosted inference private. When a hosted provider is used, some information is sent to that provider. Side Quest documents these data flows and provides configuration options for developers who want more control.

Open innovation makes this project adaptable, inspectable, and easier for the community to build upon.

My Agent Session

The agent pipeline, provider fallback logic, and quest-generation implementation are available in the repository.

Explore the implementation: Side Quest on GitHub

Prize Categories

This project is submitted for the Hacktoberfest Open-Source AI Challenge — Week 1: Touch Grass.

Additional partner prize categories should be listed here only if Side Quest meets their specific eligibility requirements.

---

Built with open-source AI, a little curiosity, and the belief that the best quest sometimes starts by putting your phone down. 🌿