Touch Grass

Inspiration

Screen-time limits are easy to ignore, and they never give you anything to do instead. We wanted a nudge people would actually enjoy, followed by something better to do. So we built Spike, a sarcastic little cactus who chirps you off your phone like a friend ribbing you, then hands you three ideas that fit your life.

What it does

Touch Grass is an AI companion that runs a simple loop: Catch → Chirp → Redirect → Reward.

Catch: You clock in for a scrolling session with an in-app timer, and an iPhone Shortcut can catch you the moment you open Instagram. When time's up, Spike steps in.
Chirp: Spike roasts you using your own to-do list. You pick the spice level, from Mild to Extra Spiky. Every level stays PG, and the teasing is about the habit, never the person.
Redirect: You get three activity ideas matched to your interests, budget, and free time. Each comes with a 5-minute starter step so you can actually begin. Reroll or filter until one fits.
Reward: When you finish, you tell Spike how long it really took. The time you reclaimed becomes real-life XP (1 minute = 1 XP), printed on a "dopamine receipt" that shows your time against the plan.

You can also manage to-dos, snooze Spike for a while, or stop the chirping for the day.

How we built it

Frontend: A mobile-first React + TypeScript app built with Vite and styled with Tailwind CSS. It includes the chat-bubble chirps, swipeable activity cards, iPhone-style scroll wheels for logging time, and a paper-receipt reward screen.
Backend: A FastAPI (Python) service that handles accounts, to-dos, check-ins, and XP. Passwords are hashed with bcrypt and sessions use JWTs.
Database: Serverless Postgres on Neon, with versioned SQL migrations.
AI: The Google Gemini API writes the chirps and the activity ideas. Calls go through our backend so the API key never reaches the browser, and every AI call has a hand-written fallback so the app still works when the model is slow or unreachable.
iPhone integration: An Apple Shortcuts automation asks our /should-interrupt endpoint whether Spike should step in, and our /snooze endpoint lets the snooze button pause it.
Deployment: The frontend is on Vercel and the API is on Render.

Challenges we ran into

Detecting scrolling without a native app. iOS only exposes app usage through Screen Time APIs that need special Apple entitlements, so we used a Shortcuts automation and an in-app scrolling timer instead.
Keeping the roasts funny but kind. We enforced PG-only, tease-the-habit rules in the prompts, in checks on the output, and in written fallback chirps.
Making AI feel instant. We prepared the next "time's up" chirp ahead of time, so the interrupt appears immediately.
Wiring everything together under time pressure. Getting the frontend, API, and database talking across three hosts (CORS, environment variables, cold starts) took real debugging.

Accomplishments we're proud of

A full loop that works end to end on a real phone, from the catch to a receipt with real XP saved to a database.
A character people want to interact with, instead of a settings screen they ignore.
A reward system tied to real time spent offline.

What we learned

Replacing a bad habit works better than blocking it. People respond to a nudge they enjoy plus an easy next step.
Every AI call needs a graceful fallback, so the demo never depends on a model answering in time.
Small design details, like a receipt and scroll wheels, make an app feel like a product.

What's next for Touch Grass

Real usage detection through Apple's Screen Time APIs (FamilyControls and DeviceActivity), replacing the Shortcut stand-in.
Content-aware suggestions that match what you were scrolling (cooking videos lead to cooking, gym content leads to a workout).
Streaks, weekly recaps, and friends to compete with on XP.
Making scrolling less appealing by shifting the phone's display, such as grayscale.