Download

The project

The shared counter

The whole workstation, not just the browser; no per-person data; a public method with its bounds.

Installing CarboneWatch exposes nothing about you — except, if you turn it on, the energy total of your AI use

The app observes how much data is exchanged with known AI services — never what you write, nor the sites you visit. It decrypts nothing, installs no certificate, requests no administrator rights, and opens no port: nothing can connect to it from outside. No account to create, no telemetry.

Your figures stay with you, in a file on your machine. Sharing to the total below is off by default; if you turn it on, all that leaves are hourly energy totals per service, under a random identifier renewed every month. Never any content, visited address, machine name or IP address. Sharing stops with one button, and erasure of what was already sent is requested in writing.

CarboneWatch is written by an independent developer, with a single intention: to make the carbon footprint of AI use visible, since it shows up nowhere else. Nothing being collected, there is nothing to sell — and the figure shown here is only worth anything if it is honest.

Legal notice and privacy →

Sum of the estimates voluntarily reported by CarboneWatch installations that turned sharing on. This is an aggregated order of magnitude, not a measurement: every figure inherits the uncertainty of the estimation model described below.

Install CarboneWatch on this machine →

An estimate, not a meter: the method is public and identical on every machine. This total is a floor — it only counts the machines that contribute, and no figure is tied to a person.

Accounts and access

Period

Nothing is read live. Usage is aggregated by the hour on each machine, then uploaded every fifteen minutes — never connection by connection.

CO₂e from online services —
machines that measured over the period —
median per machine per day —

Take away Export (CSV) Inventory lines (CSV) Reporting note

What that amounts to

The total above, converted into familiar things. Every reference is named with its assumption: "a petrol car" means nothing — a city car and a sports car are a factor of three apart.

Energy

    Carbon

      Orders of magnitude, per passenger and per kilometre for transport. They feed no calculation on this page: they only help read it. Occupancy changes everything — a half-empty ship doubles each passenger's share, a packed train halves it.

      By day

      Grams of CO₂e. Amber for online services, green for local models.

      • online services
      • local models

      By kind of use

      What the energy is spent on, rather than who profits from it. Five categories cross the anonymity threshold far sooner than any single service: this is where the structure of use shows up first.

      By service

      A service only appears once 5 distinct installations have reported it: below that, the page would tell you about a handful of people's usage.

      Where the energy is spent

      The electricity zone each installation declared in its settings — six zones, nothing finer. This is not geolocation: no IP address is read, and nothing is inferred from where an upload comes from.

      Local models

      Counted separately: their energy is the host machine's, which nothing makes comparable from one computer to the next.

      —

      Methodology

      Online services are billed on bytes received, capped at a model's maximum generation rate and floored to discard protocol noise — and, for services whose ratio has been measured, on bytes sent, which carry the context the model re-reads. Local models are billed on inference time. Energy is converted to CO₂e using the region where each service computes.

      The most uncertain parameter remains the ratio between bytes on the wire and tokens. Factors cannot be changed: every machine computes with those of the published method, the same for everyone, and the totals above are made with them.

      What is reported, and what is not

      Sharing is off by default and turned on explicitly in the settings. A machine that turns it on sends, per hour and per service: estimated energy, CO₂e, estimated tokens produced and read, service type, electricity grid code and application version.

      Never reported: IP address, machine name, process name, detected application, visited address, content. No account. The installation identifier derives from a secret drawn at random on the machine and changes every month; nothing ties it to a person. The application displays these identifiers, with which erasure can be requested in writing.

      Installing CarboneWatch exposes nothing about you — except, if you turn it on, the energy total of your AI use

      Your figures stay with you, in a file on your machine. Sharing to the total below is off by default; if you turn it on, all that leaves are hourly energy totals per service, under a random identifier renewed every month. Never any content, visited address, machine name or IP address. Sharing stops with one button, and erasure of what was already sent is requested in writing.

      CarboneWatch is written by an independent developer, with a single intention: to make the carbon footprint of AI use visible, since it shows up nowhere else. Nothing being collected, there is nothing to sell — and the figure shown here is only worth anything if it is honest.