Touchgrass.local — An AI Coach That Gets You Outside

Aniruddha Pandey — Tue, 06 Oct 2026 20:36:26 +0000

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

What I Built

Touchgrass.local is a small AI-powered browser app that encourages people to step away from their screens and spend some time outdoors.

Users describe how their day has gone, and an open-weight language model running directly in the browser gives them one small outdoor mission along with a friendly nudge.

Once the mission is completed, the user can mark it as done and watch a field of CSS grass grow. The app also tracks completed missions and streaks using localStorage.

The project is designed to be simple, private, lightweight, and useful for students and anyone who spends too much time in front of a screen.

Demo

🚀 Live Demo:
[Add your deployed GitHub Pages link here]

🤖 Claude Artifact:
https://claude.ai/artifact/RaTCVWFjQi3FrfYGTA213Z

Code

💻 GitHub Repository:
https://github.com/aniruddhapandey01/Touchgrass.local

The repository contains the complete source code and is built using plain HTML, CSS, and JavaScript.

How I Built It

Touchgrass.local uses WebLLM by MLC AI to run an open-weight language model directly inside the browser using WebGPU.

The project supports models including:

Gemma 2 2B
Qwen 2.5 1.5B
Llama 3.2 1B

The application is built with:

HTML — page structure
CSS — UI, themes, animations, and growing grass
JavaScript — model loading, prompting, mission generation, streaks, and local storage
WebLLM — in-browser AI inference

There is no backend and no API key. After the model is downloaded, it is cached by the browser and can work offline. User input stays on the device instead of being sent to a server.

Why Does Open Innovation Matter?

Open innovation makes it possible to build AI applications without depending entirely on closed APIs.

For Touchgrass.local, using open-weight models and WebLLM makes the project more private, accessible, customizable, and transparent. The model runs locally in the browser, so there is no API bill or backend required.

It also makes experimentation easier. Users can switch between different supported models and compare their behaviour.

Most importantly, the project can be shared openly so other developers can learn from it, modify it, add new missions, improve the UI, or contribute new ideas.

My Agent Session

🤖 Claude Artifact / Agent Session:
https://claude.ai/artifact/RaTCVWFjQi3FrfYGTA213Z

Prize Categories
Hacktoberfest Open-Source AI Challenge — Week 1: Touch Grass
Open-Source AI
Local / In-Browser AI
Open-Weight Models
Final Thoughts

Touchgrass.local started with a simple idea:

What if AI could encourage us to spend less time using technology?

Instead of making another app that keeps users on their screens, I wanted to build something that uses AI to give people a reason to step away from the screen and do something in the real world.

BuddyBoost Pro: A Study Productivity App Built for a Friend

Aniruddha Pandey — Sun, 04 Oct 2026 20:10:16 +0000

BuddyBoost Pro — Hacktoberfest Weekend Challenge Submission
This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built
I built BuddyBoost Pro for my friend — the one who always says "I'll start in five minutes" and then doesn't.

It's a study productivity web app that removes every excuse to delay starting. The moment you type a goal, it instantly generates a step-by-step study plan and drops you into a fast Pomodoro timer, so "getting started" takes seconds instead of willpower. It combines:

📅 Study Planner — instant, personalized session plans based on your goal, subject, time, energy, and preferred coaching vibe (gentle, structured, or playful)
⏱️ Pomodoro Timer — a live circular focus timer with phase tracking, +1/+5 min extensions, skip, and an audio chime (plus a hidden Turbo Mode for testing)
🎧 Ambient Sounds — gentle rain, ocean waves, cozy brown noise, and a 40Hz gamma binaural focus beat, all synthesized live in the browser
🎴 Flashcards — 3D flip cards for active recall, with "Needs Review" / "Mastered" tracking and instant deck generation
📊 Analytics & Streaks — total focus minutes, sessions completed, cards mastered, and a daily 🔥 streak to keep motivation up
♡ Better Together — an accountability card that writes a personal check-in message you can copy and send straight to a friend
🌗 Dark / Light Mode
Everything runs instantly in the browser — no logins, no waiting, no setup — because the whole point was lowering the friction between my friend and their first focused minute.

Demo
🔗 Live demo: https://aniruddhapandey01.github.io/buddyboost/

Open it, type a goal like "Understand arrays in C", hit Build my study plan, and press Start Session. Try the ambient sounds and flip a few flashcards too.

Code
The full source (HTML, CSS, JavaScript, plus a README with setup instructions) is ready for GitHub:

📁 Repository: https://github.com/aniruddhapandey01/buddyboost

git clone https://github.com/aniruddhapandey01/buddyboost.git
cd buddyboost
bun install # or npm install
bun run dev # opens at http://localhost:8080
Tech stack: TanStack Start (React 19), TanStack Router + TanStack Query, Vite 7, Tailwind CSS v4, and shadcn/ui-style components built on Radix UI.

How I Built It
I built BuddyBoost Pro with Lovable, an open AI agent harness — I described what I wanted in plain language, and the agent generated, structured, and refined the full codebase with me reviewing and iterating along the way.

The app itself runs on a fully open-source stack:

TanStack Start + React 19 for the app shell and routing
Tailwind CSS v4 for the themed dark/light design system
Web Audio API for synthesizing all four ambient soundscapes from scratch — no audio files, no cloud calls
A lightweight rule-based plan-generation engine that runs 100% client-side, which is why plans appear instantly instead of waiting on a server round-trip
localStorage for saving sessions, streaks, and flashcard progress privately on the user's own device
I deliberately kept the generation client-side rather than calling a closed AI API: it makes the app free to run, instant to respond, private, and fully auditable — anyone can read the code and see exactly how their study plan is built.

Why Does Open Innovation Matter?
Open innovation is the reason a student project like this can exist at all. Every layer of BuddyBoost Pro — the framework, the router, the build tool, the styling system, the UI primitives — is open source that I could use for free, inspect, and build on.

That mattered in three concrete ways:

No paywall between my friend and focus. If the plan generator depended on a paid closed API, every study session would cost money or need my API keys. Because it's open, transparent, client-side code, the app runs free forever, for anyone.
Privacy by architecture, not by promise. Progress is saved in the user's own browser. With open code, anyone can verify that — no trusting a black box with study habits.
Anyone can remix it. A classmate could fork it, add spaced repetition, translate it, or port it to a different stack. Closed APIs lock ideas behind terms of service; open code lets ideas travel.
My Agent Session
I built this project in a Lovable agent session — the agent scaffolded the app, wired up the timer, sounds, flashcards and analytics, and iterated with me on packaging it for GitHub (source ZIP, standalone script.js, and README.md).

Prize Categories
Build for a Friend — BuddyBoost Pro was designed around one real person's procrastination, with accountability features (streaks, friend check-in messages, tiny first wins) built for exactly that struggle
Thanks for participating! 🎃

DEV Community: Aniruddha Pandey

Touchgrass.local — An AI Coach That Gets You Outside

BuddyBoost Pro: A Study Productivity App Built for a Friend