Run your personal AI agents on the old devices in your drawer instead of new hardware that adds to CO₂ emissions.
A solo project built at Hack the Valley 2026.
An always-on agent needs a machine that stays on. Buying one costs money and adds manufacturing emissions. Running it on your daily phone wears out its battery.
CO₂LAB uses the devices you already retired instead: any Android 11+ phone and any Windows, macOS or Linux computer. A hub plans each request with Gemini and sends the work to whichever old device can do it, with a cloud worker as backup. The dashboard shows each request's CO₂e, and the Lab tests cheaper ways to run the agents.
- Each device connects to the hub and reports its skills (apps, browser, Steam games), battery and temperature.
- Gemini reads those reports and splits the request into tasks.
- The scheduler gives each task to a device that can do it. Hot or offline devices are skipped.
- The phone works inside apps through the Accessibility Service. The laptop and cloud worker drive a browser with Playwright.
- When a step needs judgement, the device sends its screen to the hub and Gemini picks the next action.
- Gemini combines what the devices found into one answer. Every call is logged in Tiger Data.
- Flash plans requests and writes answers. If it is out of quota, the hub tries the older Flash models in
GEMINI_MODEL_FLASH_BACKUPS, then Flash-Lite. - Flash-Lite steers the devices and reads screens. The Lab found it as accurate as Flash on the deal screens at a quarter of the cost.
- All calls use one API key and go through the hub. Devices never hold a key.
- App agent on Android: opens Tim Hortons, McDonald's and Popeyes through the Accessibility Service and screenshots their offers.
- Web agent: Playwright browses sites such as RedFlagDeals on a laptop or the cloud worker.
- General requests: the planner reads what each device has (installed apps, Steam games, browser) and picks tasks to match.
- Health-aware scheduling: work moves off a device that goes offline. Devices pause at 90 °C (laptop CPU) or 42 °C (phone battery).
- Burn meter: each request shows its CO₂e, the old devices that worked on it, and the AI cost.
- Lab, Impact: every Gemini call from Tiger Data, broken down by step and device. Click a request to see only its calls.
- Lab, Research: cites papers from Papiers, runs trials on labelled screenshots and saves findings. You can type your own idea to test.
- Compute jobs: Steam review analysis split into batches across devices.
Android phone (Kotlin) Laptop / cloud worker (Node + Playwright)
Accessibility Service browser agent, review batches
\ /
\ WebSocket (zod) /
v v
Hub (Node, Fastify, ws) on Vultr behind Caddy
planner, scheduler, health engine, research loop
| | |
Gemini API Tiger Data Papiers CLI
(Postgres + Timescale) (read-only)
|
v
Dashboard (React PWA): devices, agent runs, live screens, Lab
- Frontend: React 19, Vite, vite-plugin-pwa
- Backend: Node 24, TypeScript, Fastify, ws, zod,
@google/genai, pg - Devices: Kotlin Android app (OkHttp), Node worker with Playwright and Transformers.js
- Services: Gemini API, Tiger Data, Papiers, Vultr with Caddy and systemd
- Node.js 22 or newer
- A Gemini API key from Google AI Studio
- Optional: a Tiger Data (Postgres) database. Without
DATABASE_URLthe hub keeps state in memory. - Optional: a Papiers token for the research loop
- Optional: Android Studio to build the phone app
git clone https://github.com/turtle-geek/co2lab.git
cd co2lab
cp .env.example .env # then fill in your keys
npm install
npm run dev -w hub
npm run dev -w dashboardOpen http://localhost:5173 and log in with your OWNER_PASSPHRASE. To try the dashboard without devices, open http://localhost:5173/?mock=1 (passphrase demo).
Add a laptop from the dashboard with Add device, or from the repo:
npm run pair -w worker-web
npm run start -w worker-webBuild the Android app with Android Studio, or from android/ with ./gradlew assembleDebug. Pair it with the QR code from Add device.
| Variable | What it is for |
|---|---|
GEMINI_API_KEY |
The one Gemini key the hub uses |
GEMINI_MODEL_LITE, GEMINI_MODEL_FLASH, GEMINI_MODEL_PRO |
Model IDs for each tier |
GEMINI_MODEL_FLASH_BACKUPS |
Older Flash models to try, in order, when Flash is out of quota |
OWNER_PASSPHRASE |
Dashboard login |
SESSION_SECRET |
Signs the login cookie (32+ characters) |
DATABASE_URL, DATABASE_PASSWORD |
Tiger Data connection. Leave empty to run in memory |
PAPIERS_TOKEN |
Read access to your Papiers library for Lab research |
HUB_PUBLIC_URL, HUB_PORT |
Public address and port of the hub |
WORKER_HUB_URL |
Hub address for a worker run from the repo |
VULTR_SERVER_IP, HUB_DOMAIN |
Used by deploy/ scripts. HUB_DOMAIN is the public name, for example co2lab.work |
bash setup-server.sh <hub-domain> # once, as root on a fresh Ubuntu 24.04 server
bash deploy/deploy.sh # ships pushed code; add --env to upload .env| Method | Endpoint | What it does |
|---|---|---|
POST |
/api/login |
Owner login, sets the session cookie |
POST |
/api/device-code |
A computer asks to join |
GET |
/api/impact |
Gemini usage by step, device and request |
GET |
/api/research |
Lab captures and findings |
GET |
/api/devices/:id/telemetry |
A device's battery and temperature history |
GET |
/api/health |
Hub status |
WS |
/ws |
Devices and dashboards. Messages are defined in shared/src/protocol.ts |
hub/ Node hub: planner, scheduler, Gemini, research loop, API
dashboard/ React PWA
worker-web/ Laptop and cloud worker (Playwright, review batches)
android/ Kotlin phone agent
shared/ zod message contract used by every part
db/ Tiger Data schema
deploy/ Vultr setup, Caddy and systemd files