Clementine

Transform idle home devices into a private distributed compute cluster.

Clementine lets you submit a job — by text or voice — and automatically splits it into parallel subtasks distributed across every device on your local network. No cloud, no fees, no data leaving your home.

---

How it works

1. Host opens the dashboard and starts a session
2. Workers join by scanning a QR code or visiting the session URL on any device
3. Submit a job — Clementine uses an LLM to decompose it into parallel subtasks
4. Tasks are distributed to idle workers and executed in real-time in the browser
5. Results are aggregated and synthesized into a final summary by Gemini

---

Stack

Layer	Tech
Framework	Next.js 16 (custom Node.js server)
Real-time	Socket.IO
LLM Orchestration	LangGraph + LangChain
Task Decomposition	K2 (k2think.ai)
Result Synthesis	Gemini 2.5 Flash
Animations	Framer Motion, Canvas API
Database	MongoDB (job history)
Styling	Tailwind CSS v4

---

Getting started

Prerequisites

• Node.js 18+

Install

git clone https://github.com/priyam-k/Clementine.git
cd Clementine
npm install

Environment variables

Create .env.local:

# Task decomposition (required)
K2_API_KEY=your_k2think_api_key
K2_BASE_URL=https://api.k2think.ai/v1
K2_MODEL=MBZUAI-IFM/K2-Think-v2

# Result synthesis (required)
GEMINI_API_KEY=your_gemini_api_key

# Anthropic fallback for decomposition (optional)
ANTHROPIC_API_KEY=

# Job history persistence (optional)
MONGODB_URI=

# Set this in production to your public domain
# RAILWAY_PUBLIC_DOMAIN=your-app.up.railway.app

Run

npm run dev

Open http://localhost:3000 as the host. Workers join at http://<your-local-ip>:3000/join on any device on the same network.

---

Pages

Route	Description
/	Landing page
/host	Host dashboard — dispatch jobs, monitor workers and tasks live
/host/tasks	Task inspector — per-task input/output breakdown with job process dump
/host/workers	Worker list with device telemetry
/join	Worker join page — scan QR or enter session code

---

Voice commands

The host dashboard supports browser speech recognition:

• Click the mic button in the Voice Command card
• Say "Hey Clementine, <your task>"
• Edit if needed, then dispatch

Requires Chrome or Edge (webkitSpeechRecognition).

---

Deploying

Clementine requires a persistent Node.js process — Socket.IO needs long-lived connections and won't work on Vercel or other serverless platforms.

Recommended: Railway

1. Connect your GitHub repo on Railway
2. Add all env vars in the Variables tab
3. Set RAILWAY_PUBLIC_DOMAIN to your generated *.up.railway.app domain
4. Click Generate Domain in the Railway service settings for a public URL

---

Project structure

server.ts                    Custom HTTP + Socket.IO + Next.js server entry point

src/
  app/                       Next.js pages + API routes
  components/
    host/                    Dashboard UI (JobCard, WorkerGrid, TaskInspectOverlay, ...)
    join/                    Worker UI
    landing/                 Landing page sections + canvas visualizations
    ui/                      Primitives (ProgressBar, Badge)
  server/                    Server-side logic
    decomposer.ts            Job → subtask splitting
    scheduler.ts             Task → worker assignment
    reducer.ts               Output aggregation
    graph-nodes.ts           LangGraph node implementations
    socket-handlers.ts       All Socket.IO events
  hooks/
    useHostSession.ts        Host dashboard state
    useWorkerSession.ts      Worker state + task execution
  lib/
    shared-types.ts          Wire types shared by server and client
    compute-executors.ts     Real JS computation implementations

---

Built at YHack 2026