Jums — The AI Life Operating System
One conversation. Every platform. Entirely yours.
Inspiration
We looked at how people actually live their lives in 2026 and saw a brutal disconnect. Everyone has a dozen apps for productivity, health, messaging, journaling, reminders, and goals — yet nobody feels in control. The tools that were supposed to free us have fragmented our attention across platforms, notifications, and dashboards that demand constant babysitting.
We asked a simple question: What if you only had to do the human part — set your goals, make decisions, reflect — and an intelligent agent handled everything else?
Not a chatbot. Not another dashboard. A genuine autonomous life operating system that knows your goals, tracks your health, manages your schedule, learns your preferences over time, and proactively works on your behalf — across every platform you already use.
The inspiration came from studying the OpenClaw open-source agentic framework — a battle-tested multi-channel AI gateway that connects LLMs to WhatsApp, Telegram, Slack, Discord, and 15+ messaging platforms through a single WebSocket control plane. We saw the raw power of an agent that can use tools, schedule cron jobs, maintain vector memory, spawn sub-agents, and operate across channels. We thought: what if we took that agentic architecture and pointed it at the most important product category in the world — your life?
That's Jumns. An agent-first life assistant where the conversation IS the interface. One thread. No tabs. No dashboards. Just you and an AI that actually works for you.
What It Does
Jumns is a personal AI life assistant that runs as a native app connected to a serverless backend. The agent doesn't wait for you to ask — it proactively manages your day through rich interactive cards pushed directly into a single conversation stream.
Morning: Jumns pushes a Daily Briefing card — weather, your goals for the day, scheduled tasks, a progress ring showing how you're tracking. You tap "Looks good" or adjust on the fly.
Throughout the day: Reminder cards surface at the right time. Goal check-in cards track your streaks and progress. The agent dynamically adjusts priorities based on what's actually happening — a meeting ran long, you skipped your run, a deadline moved. Jumns recalibrates in real time.
Evening: A Journal Prompt card invites you to reflect. Mood tracking. Day summary. Everything saved to persistent vector memory so the agent builds a deeper understanding of you over weeks and months.
Health: Continuous vitals tracking with OpenEHR integration on the roadmap — steps, sleep, heart rate, trends over time. Not just numbers, but an AI that interprets them: "Your sleep has dropped 40 minutes this week. Your afternoon energy dips correlate. Want me to adjust your evening routine?"
The Toolkit: Jumns is extensible. MCP servers for external data (weather, calendar, Spotify). A2A agent delegation for complex tasks (a research agent, a coding assistant). Context sources from Google Drive, local files, or any knowledge base. Skills that plug in like apps within the agent.
Voice Mode: Full-screen voice interaction with a living orb visualization. Hands-free conversations with your agent. Ask anything, get things done, no screen required.
How We Built It
Jumns is built on a modern, cost-efficient stack designed to scale from zero to millions:
Frontend — React 18 + TypeScript + Tailwind CSS v4. Mobile-first responsive design with Framer Motion animations. shadcn/ui component library built on Radix UI primitives. White Canvas aesthetic with purple/pink accents. The app is a thin, beautiful UI client — all intelligence lives on the backend.
Backend — Node.js + Express 5 + TypeScript. RESTful API with 21 AI-powered tools. Google Gemini 3 Pro as the primary model (fast, capable, excellent tool use). PostgreSQL with pgvector for persistent vector memory and embeddings.
Agent Architecture — Inspired by OpenClaw's embedded runner pattern: dynamic system prompt assembly per turn, tool-calling loop with max iteration guards, context compaction when history exceeds the model window, and a skills framework for pluggable tool collections. The agent doesn't just respond — it plans, remembers, and acts.
Card System — Rich interactive cards rendered natively within chat messages. Daily briefings, goal check-ins, reminders, journal prompts, health snapshots — all pushed proactively by the agent. Card types are detected from actual tool call results, not text heuristics.
Memory System — Local feature-hashing embeddings (768 dimensions) stored in pgvector. Cosine similarity search for context retrieval. Auto-extraction of important facts after every conversation. Deduplication via Jaccard similarity. Categories include preferences, personal info, goals, habits, health, and relationships.
Target production stack: Kotlin + Jetpack Compose Android app, Python 3.12 + FastAPI on AWS Lambda (serverless), Strands Agents SDK, Supabase Auth, RevenueCat billing. Cost at MVP scale: ~$5-10/month total.
Challenges We Ran Into
Designing the "no-dashboard" paradigm. Every productivity app defaults to tabs, lists, and dashboards. We had to fight the instinct to add complex navigation. The breakthrough was realizing the agent IS the interface — one conversation stream with rich cards replaces an entire app ecosystem.
Card rendering in a chat stream. Mixing rich interactive UI (progress bars, goal trackers, checklists) with a scrolling message list is technically challenging. We built a card detection system that maps tool call results to card types, with graceful fallback to plain text.
Proactive agent behavior without being annoying. The agent needs to push cards at the right time without becoming notification spam. The agent decides whether to surface information or stay quiet — a delicate balance between helpful and intrusive.
Making the agent feel alive, not robotic. The thinking indicators, the personalized name and personality, the voice mode with visual feedback, the contextual greetings — these small details transform a chat interface into something that feels like a living companion.
Persistent memory that actually works. Building a vector memory system that auto-extracts facts, deduplicates entries, and retrieves relevant context per-turn required careful engineering of embeddings, similarity thresholds, and category management.
Accomplishments We're Proud Of
Agent-first architecture that actually works. We proved that a single conversation stream with proactive cards can replace an entire productivity app suite. No dashboards, no tab-switching, no cognitive overhead. The agent handles the complexity; the human handles the decisions.
Fully personalized experience. From the moment you name your agent and choose its personality, every screen reflects your choices. The agent's name appears in headers, placeholders, thinking indicators, and chat bubbles. It feels like YOUR assistant, not a generic chatbot.
21 functional AI tools. Goal CRUD, task management, reminders, daily summaries, progress analysis, smart suggestions, web search, and long-term memory — all working through natural conversation.
Production-grade agent architecture. Dynamic system prompt assembly, tool-calling loops with iteration guards, context compaction, vector memory with pgvector, and proactive scheduling. This isn't a wrapper around an API — it's a full agentic runtime.
Voice mode with Web Speech API. Real speech-to-text that sends transcriptions to the AI agent, with visual feedback and error handling for unsupported browsers.
Cost efficiency that enables a real business. Free tier potential, Pro at $9.99/month. Backend costs under $10/month at development scale. RevenueCat handles billing complexity. The unit economics work from day one.
What We Learned
The future of apps is agents, not interfaces. Users don't want more screens to manage. They want outcomes. An agent that proactively manages your day, adjusts to changing circumstances, and learns from your patterns is fundamentally different from a todo app with AI sprinkled on top.
Conversation is the most natural interface for life management. People already manage their lives through messaging — texting reminders to themselves, voice-noting ideas, chatting with friends about plans. Jumns meets them where they already are.
Proactive > Reactive. The biggest unlock isn't answering questions — it's anticipating needs. A morning briefing you didn't ask for. A reminder that adjusts because your meeting ran late. A journal prompt that references what actually happened today. The agent that acts before you ask is the agent you can't live without.
Health data without context is noise. Steps, sleep, heart rate — these numbers mean nothing in isolation. An AI that correlates your sleep drop with your afternoon energy dip and suggests a concrete change? That's the product.
The agentic framework patterns are battle-tested and portable. Gateway-centric architecture, session management, tool policies, vector memory, skills framework — these patterns translate directly from enterprise AI systems to a personal life assistant. The architecture is the moat.
What's Next for Jumns
Phase 1 — Ship the Native MVP (Now)
Launch the Android app with core agent conversation, daily briefings, goal tracking, reminders, journaling, and health snapshots. Kotlin + Jetpack Compose frontend. Python + FastAPI + Strands Agents SDK backend on AWS Lambda. Supabase Auth, RevenueCat billing, Gemini 2.5 Flash.
Phase 2 — Preventive Health with OpenEHR (Q3 2026)
Continuous vitals tracking through OpenEHR-compliant integrations. Wearable device connections (Apple Health, Google Health Connect, Fitbit, Garmin). The agent doesn't just display metrics — it builds a longitudinal health model, detects patterns, flags anomalies early, and suggests preventive interventions. "Your resting heart rate has trended up 8 BPM over 3 weeks. Combined with your reported stress levels, I'd recommend scheduling a check-up."
Phase 3 — Dynamic Goal Intelligence (Q4 2026)
Goals that adapt in real time. The agent analyzes your behavior patterns, environmental factors, and progress trajectories to dynamically adjust targets, timelines, and strategies. Missed your morning run three days in a row? The agent doesn't just nag — it proposes shifting to evening runs based on your calendar, or suggests a 15-minute home workout as a lower-friction alternative. Live situation analysis with continuous recalibration.
Phase 4 — Multi-Platform Expansion (2027)
iOS app. Web client. Multi-channel messaging integration (WhatsApp, Telegram, Slack) — leveraging the OpenClaw gateway patterns. Your agent accessible from any device, any platform, any conversation. One agent, everywhere.
Phase 5 — The Agent Marketplace (2027)
Open the skills and A2A agent framework to third-party developers. Specialized agents for fitness coaching, financial planning, meal prep, language learning, therapy support — all plugging into the Jums runtime. The App Store moment for personal AI agents.
Phase 6 — Enterprise & Teams (2028)
Team goal alignment. Shared agents for project coordination. Organization-wide health and wellness programs with privacy-preserving analytics. The personal life assistant becomes the workplace operating system.
The Vision
Humans should only need to do the highest-level work — setting goals, making decisions, reflecting on what matters. Everything else — tracking, scheduling, adjusting, reminding, analyzing, coordinating — should be handled by an agent that knows you deeply and works tirelessly on your behalf.
Jumns is the first step toward that future.
We're not building another productivity app. We're building the operating system for human potential.
Log in or sign up for Devpost to join the conversation.