Hackbitz

One companion per goal. They talk to each other so you don't have to.

Most habit trackers just count streaks. Hackbitz gives each of your goals its own AI companion that actually reads your logs, notices patterns across your life, and reaches out through Telegram when something needs attention. Skip the gym three days in a row while your sleep is great? Your fitness companion knows, and it also knows your sleep companion isn't worried — so it nudges you gently instead of panicking.

---

What it does

You tell Hackbitz what you're working on in plain English:

"Go to the gym 4 times a week" "Sleep 8 hours, be in bed by 11pm" "Keep weekly spending under $150"

Each goal gets its own companion (fitness, sleep, money, social, or deadline-based). You share updates whenever — through the web app or by texting a Telegram bot. The system figures out which goal your message belongs to and routes it automatically.

Every 15 minutes (and reactively when you log something), all companions run in parallel:

• Read your recent logs
• Check what the other companions are thinking
• Search semantic memory for cross-goal context
• Decide: watch, nudge, call out, or escalate

Then a coordinator steps in. If multiple companions are concerned, it picks the ONE thing that matters most (sleep > deadlines > fitness > money) and sends you a single Telegram message — not five. If nothing's wrong, it stays quiet.

When a companion does reach out, it also pulls in a relevant article, podcast, or tool from Exa AI — so you get something actionable, not just "hey, do better."

---

How it works

You (web app / Telegram)
  │
  ├─ "Skipped gym again, feeling lazy"
  │     │
  │     ├─ LLM classifies → fitness goal
  │     ├─ Saved to Supabase (user_logs)
  │     └─ Triggers agent tick
  │
  ▼
Modal Agent Network (parallel)
  │
  ├─ Fitness agent: "3 skips this week, 0/4 target" → nudge
  ├─ Sleep agent: "8h avg, consistent bedtime" → monitor (all good)
  ├─ Money agent: "$92 spent, under $150" → monitor
  └─ Social agent: "cancelled plans 3x this month" → concerned
  │
  ▼
Coordinator
  │
  ├─ 2 agents want to intervene (fitness + social)
  ├─ LLM resolves: "Focus on fitness. Social can wait."
  ├─ Pulls Exa article: "10-min home workout for zero motivation days"
  └─ Sends ONE Telegram message with everything

The companions also talk to each other through shared state. If your sleep companion is escalating, your fitness companion backs off — because pushing someone to work out when they're sleep-deprived doesn't help.

---

Telegram bot

Text @hackbitz_bot on Telegram. It's the main way companions reach you.

Commands:

• /status — quick health check across all goals (no LLM, instant)
• /confused — "just tell me the ONE thing I should do right now"
• /plan — today's top 3 priorities based on current state
• /help — what the bot can do

Or just text it anything: "slept terribly", "spent $40 on lunch", "went for a run today" — it classifies, saves, and triggers the right companion.

---

Tech stack

Layer	What	Why
Frontend	Next.js 15, React 19, Tailwind v4, shadcn/ui	fast to build, looks clean
Database	Supabase (Postgres)	real-time, good free tier, instant API
Backend	Modal (Python)	cron scheduling, parallel .map() over goals, zero infra
LLM	Groq API (llama-3.1-8b for agents, llama-3.3-70b for coordinator)	free tier, ~200ms latency, no GPU needed
Memory	Supermemory	semantic search across observations, cross-goal context
Content	Exa AI	finds relevant articles/podcasts/tools when companions nudge
Delivery	Telegram Bot API	instant, personal, no app to install

No GPU. No vLLM. Modal runs CPU-only functions; Groq handles all inference. The whole thing runs on free tiers.

---

Project structure

├── frontend/                    # Next.js web app
│   ├── app/
│   │   ├── api/goals/          # CRUD + LLM classification
│   │   ├── api/logs/           # Activity logs + auto-routing
│   │   ├── api/agent-messages/ # Companion messages
│   │   ├── api/agent-states/   # Agent internal states
│   │   └── api/trigger-tick/   # Manual agent trigger
│   ├── components/             # GoalForm, ActivityLog, AgentChat, AgentMemory
│   └── lib/                    # Supabase client, Groq client
│
├── backend/
│   ├── modal_app/
│   │   ├── app.py              # Modal entrypoint: cron, parallel agents, coordinator
│   │   ├── coordinator.py      # Cross-agent conflict resolution, Telegram delivery
│   │   └── agents/             # fitness, sleep, money, social, short_lived
│   │       └── base.py         # BaseAgent: logs, memory, peer states, LLM assess
│   └── shared/
│       ├── supabase_client.py  # DB operations
│       ├── supermemory_client.py
│       ├── exa_client.py       # Content suggestions with rotating "flavors"
│       └── telegram_client.py
│
├── supabase/migrations/        # SQL schema
└── docs/openapi.yaml           # API documentation (Swagger/OpenAPI 3.0)

---

Setup

1. Get your keys

You need accounts on: Supabase, Groq, Modal, Supermemory

Exa AI and Telegram are optional — the system works without them, just no content suggestions or Telegram delivery.

2. Create the database

In your Supabase SQL editor, run the contents of supabase/migrations/001_initial_schema.sql. Then disable RLS for Phase 1:

ALTER TABLE goals DISABLE ROW LEVEL SECURITY;
ALTER TABLE user_logs DISABLE ROW LEVEL SECURITY;
ALTER TABLE agent_states DISABLE ROW LEVEL SECURITY;
ALTER TABLE interventions DISABLE ROW LEVEL SECURITY;
ALTER TABLE agent_messages DISABLE ROW LEVEL SECURITY;

3. Backend

pip install modal
modal token new

# Create the secret with your keys
modal secret create lifeos-secrets \
  GROQ_API_KEY=your_key \
  SUPABASE_URL=https://your-project.supabase.co \
  SUPABASE_SERVICE_KEY=your_service_key \
  SUPERMEMORY_API_KEY=your_key

# Deploy
modal deploy backend/modal_app/app.py

4. Frontend

Create frontend/.env.local:

NEXT_PUBLIC_SUPABASE_URL=https://your-project.supabase.co
NEXT_PUBLIC_SUPABASE_ANON_KEY=your_anon_key
GROQ_API_KEY=your_key
MODAL_TRIGGER_URL=https://your-workspace--lifeos-agents-trigger-tick.modal.run

cd frontend
npm install
npm run dev

Open http://localhost:3000 — add a goal, share an update, hit "Check in now."

---

What makes this different

• Companions, not dashboards. You don't check a dashboard. They come to you.
• They coordinate. Multiple agents concerned? You get ONE message, not five notifications.
• Silence is a feature. If everything's fine, you hear nothing. That's intentional.
• Cross-domain awareness. Your fitness companion knows your sleep companion is worried. Agents read each other's states.
• Content, not just reminders. When a companion nudges you, it includes a relevant article, podcast, or tool — something you can actually act on.
• Dedup built in. No spam. Max one intervention per 6 hours per user. Win celebrations bypass this (because good news should always get through).

---

API docs

Full OpenAPI 3.0 spec at docs/openapi.yaml. Paste it into editor.swagger.io or import to SwaggerHub.

Method	Endpoint	What it does
GET	/api/goals	List all goals
POST	/api/goals	Create goal (LLM auto-classifies from plain English)
GET	/api/logs	Recent activity logs
POST	/api/logs	Log an update (auto-matched to goal)
GET	/api/agent-messages	Companion messages
GET	/api/agent-states	What each companion is currently thinking
POST	/api/trigger-tick	Manually run all agents + coordinator

---

Built at a hackathon. The idea: what if your goals had friends that actually talked to each other?