AI · Behavioral Tech · Mental Health · Live · Beta

Cravelock AI

An AI-powered impulse intervention app that intercepts the moment between urge and action. It stops you before you give in, not after.

Try Cravelock on Telegram

Live Beta Telegram Mini App No signup required

Designed for educational and self-management purposes. Not a replacement for professional medical care.

Timeline2026 · MVP

CategoryBehavioral AI

TypeIndependent Project

PlatformTelegram Mini App

DeploymentVercel

Status● Prototype · Active Development

Overview

What Cravelock AI does.

Cravelock AI is a Telegram Mini App that activates the moment you feel an urge. It walks you through a structured 90-second intervention, grounding you before you act, not after. The full loop runs inside Telegram with no login, no friction, no delay.

Every session is tracked: streak, resistance rate, trigger patterns, and a 7-day history, so the user can see their progress across time.

The problem

The problem no tool solves.

Most addiction and habit tools exist before or after the urge. Therapy, journals, accountability apps. None of them help during it. An urge usually peaks and fades within minutes. That window is where the decision actually happens, and it has had no dedicated tool.

Cravelock fills that window.

Why I built it

Built from lived experience.

I built Cravelock AI from personal experience with behavioral addiction. I went looking for something that would help in the moment (during the urge, not before or after) and found nothing. That absence, not a market analysis, is what created this. The design is not theoretical. Every screen, every word, every mechanic was written from the inside of that experience.

Architecture

How the intervention system is structured.

Telegram User

Telegram Mini App

Vanilla HTML · CSS · JavaScript · WebView

Telegram SDK

Haptics · Events

→

Core Differentiator

Intervention State Engine

Trigger → Hard Stop → Commit → Timer → AI → Exit

←

Timers

5–8s countdown

AI Grounding Engine

Claude

claude-opus-4-6 · Tone adaptation · 200 token cap · <2s response

Session Storage

In-memory · MVP

→

Session Manager

Telegram user ID · State scoping · No login

←

Node.js Backend

Express · Vercel

Tracking Layer

Streak · Resistance rate · Trigger log

↓

Metrics Dashboard

7-day history · Pattern visibility

How it works

The 90-second intervention loop.

Trigger

User taps "I FEEL THE URGE" inside Telegram. App asks what triggered it — an emotion or a thought/image.

Hard stop

"DO NOT FEED IT. Stay still." — A full-screen interrupt. No buttons. No way to skip. The urge has nowhere to go.

The commitment tap

"Hold it. Don't act." — A single button: I'M HOLDING. Nothing starts until the user taps it. That tap is a conscious choice made under pressure. It is the most important moment in the loop.

5–8 second timer

A countdown runs. The user holds. The urge begins to lose momentum.

AI grounding

An AI chat opens. Claude reads the user's tone and urgency in real time, then responds: firmly, steadily, or reflectively. "That urge isn't you. Let it pass."

Exit

"You didn't follow it. That's control." — Session complete. Streak updated. Trigger logged.

Session tracking

Streak, resistance rate, trigger patterns, and 7-day history, all tracked per Telegram user ID. No account required.

Demo

See the intervention loop.

What you'll see:

✓ User starts an intervention session

✓ Hard-stop screen interrupts the urge

✓ Commitment tap activates the timer

✓ Claude adapts its response in real time

✓ Session metrics update automatically

✓ Streak and trigger history are recorded

Try Cravelock on Telegram

Outcomes

Where it stands.

90s

Full intervention loop

Beta

Closed beta, active testers

User reported avoiding a relapse after using the loop

1 user reported avoiding a relapse after using the intervention loop. Beta testing ongoing via X and Telegram communities.

Open Cravelock on Telegram

Engineering challenges

What was hard to build.

The commitment mechanic

The timer only starts on the user's deliberate tap. Nothing can be skipped or auto-advanced. This required careful state management across the frontend and backend to ensure the flow couldn't be bypassed, even accidentally. The tap itself is the intervention, not the timer.

AI tone adaptation in real time

Claude reads the user's recent message length and urgency keywords in real time to decide whether to respond firmly, steadily, or reflectively. The model is capped at 200 tokens for speed. Responses need to land in under 2 seconds or the moment is lost.

Telegram Mini App constraints

No standard browser APIs, no persistent sessions by default, haptic feedback via Telegram SDK, and everything had to work inside Telegram's WebView without a login screen. The Telegram user ID replaces authentication entirely, which required rethinking how session state is scoped and protected.

Built with

The stack behind it.

Claude API (claude-opus-4-6) 200 token cap · Speed optimised

Frontend

Telegram Mini App Vanilla HTML · CSS · JavaScript Telegram SDK (haptics)

Backend & Deployment

Node.js Express.js Vercel In-memory state (MVP)

What this tells you

What I bring to a team.

Cravelock AI is the hardest product I've built. Not technically, but in terms of getting the design right. Every screen had to earn its place. The product works because of restraint: no features that don't belong in a 90-second window, no language that softens the moment, no UX that lets the user escape too easily.

Cravelock AI demonstrates designing software where timing, psychology, and interaction design matter as much as the code itself.