🦘 watts is watching your tokens GPT-4o · ~0.3g CO₂ per 1K tokens claude sonnet · ~0.1g CO₂ per 1K tokens your macbook · ~8W active load india grid · ~700g CO₂ per kWh every prompt has a price 🦘 watts is watching your tokens GPT-4o · ~0.3g CO₂ per 1K tokens claude sonnet · ~0.1g CO₂ per 1K tokens your macbook · ~8W active load india grid · ~700g CO₂ per kWh every prompt has a price

introducing carb-0n

your code has
a footprint.

Watts lives on your desktop and tracks the real carbon cost of your LLM usage, device power, and screen time. In real-time.

↓ download free see how it works
0.3g
CO₂ per 1K GPT-4o tokens
700g
CO₂ per kWh on India's grid
0
other tools track this for devs

meet watts — he has opinions about your prompts

🦘
happy
efficient session · low emissions · good prompts
😓
stressed
building up · 30 min of heavy usage
🔥
overload
you asked GPT-4 to write hello world. again.
💤
resting
device idle · emissions paused · watts approves

features

everything Watts tracks
while you ship

No dashboards to open. No forms to fill. Watts just watches and reacts.

⚡
LLM token tracking
Every API call to ChatGPT, Claude, or Gemini is counted, converted to CO₂, and added to your running total.
💻
device power monitoring
CPU, RAM, and battery drain tracked via psutil. Multiplied by your region's live grid carbon intensity.
baking
🌍
real-time grid intensity
Pulls live gCO₂/kWh data for your region from Electricity Maps API. Your number is always accurate.
baking
🧠
model swap suggester
Burning Opus on a simple task? Watts suggests a lighter model that does 80% of the job at 5% of the cost.
baking
📊
prompt efficiency score
Tracks tokens-per-outcome. Grades your prompting habits. Rewards tight, precise prompts with a better score.
baking
📖
end-of-day story
At 6pm, Watts generates a one-paragraph narrative of your day's emissions. Personal, memorable, shareable.

two data streams,
one little koala

01
intercept API calls
A lightweight SDK wrapper catches every outbound LLM call and counts tokens before returning the response.
02
measure device load
psutil polls CPU and memory every 60s, estimates wattage, multiplies by live grid carbon intensity.
03
merge into one number
Both streams combine into a rolling gCO₂ total, stored locally in SQLite. Nothing leaves your machine.
04
Watts reacts
The desktop character updates its mood state, sends nudges, and roasts you when you deserve it.
carbon0/tracker.py
# wrap your existing LLM calls
from carbon0 import track

# before
response = client.messages.create(
  model="claude-sonnet-4-6",
  messages=messages
)

# after (one line change)
response = track(client.messages.create(
  model="claude-sonnet-4-6",
  messages=messages
))

# watts gets the rest
# > session: 2.4g CO₂ · 47 prompts
# > efficiency score: B+
# > watts mood: 😓 stressed

download

get Watts on your desktop

Free, open source, runs locally. Your data never leaves your machine.

🍎
macOS
v0.2.0 · Apple Silicon + Intel
download .dmg
baking
🪟
Windows
v0.2.0 · Windows 10/11
download .exe
baking
🐧
Linux
v0.2.0 · .deb / .AppImage
download .deb
🧩
Chrome Extension
tracks AI chatbot usage in-browser
add to chrome

or install via pip: pip install carb-0n

get started in 5 minutes

installation

Install carb-0n via pip. Python 3.9+ required.

# install
pip install carb-0n

# launch Watts on your desktop
carbon0 start

wrapping LLM calls

carb-0n works by wrapping your existing API calls. Add one import and one wrapper function — that's it.

# works with OpenAI, Anthropic, Google
from carbon0 import track

response = track(your_llm_call(...))

configuration

carb-0n auto-detects your region for grid intensity. You can override in ~/.carbon0/config.toml.

[region]
zone = "IN-NO" # India North

[watts]
nudge_interval = 30 # minutes
roast_mode = true
daily_story = true

watts mood states

Watts transitions between four states based on your rolling 1-hour CO₂ total.

happy < 5g CO₂/hr · efficient session
stressed 5–15g CO₂/hr · building up
overload > 15g CO₂/hr · watts is judging you
sleeping idle detected · emissions paused

privacy

All data is stored locally in ~/.carbon0/data.db (SQLite). carb-0n makes exactly one external request: fetching live grid intensity from Electricity Maps. No prompt content is ever stored or transmitted.

why carb-0n?

Every time you ask an LLM a question, it burns energy in a data center somewhere. Most developers have zero visibility into this and that's a problem, because LLM usage is growing fast.

carb-0n was built to make that cost tangible. Not to make you feel guilty, but to make you a more intentional developer. Tighter prompts. Lighter models when they'll do. Awareness that compounds over time.

Watts is your companion in that. He's not preachy. He's just honest.

open source C++ python local-first no telemetry built for devs
🦘
hi, i'm Watts.
I live in the corner of your screen. I watch your tokens. I have opinions.

When you write a tight, efficient prompt — I'm happy.

When you ask GPT-4 to write a hello world for the third time today — we need to talk.