How much energy and carbon does your AI use? Measure it, per token.

Most teams guess their AI carbon footprint from cloud spend. CrbonFree meters every call, per model and per provider, with published factors and an uncertainty band you can defend.

  • Free plan, no card
  • Measurement uses usage metadata, not your prompts.
See what a year of OpenRouter traffic adds up to
Live · 17 Nov, 10:57 UTC

Tokens · 365 days

130B

365 days to 17 Nov 2026

CO₂e · 365 days

17.7 t

lifecycle · methodology v1.2

Providers

11

connected

Usage by provider

  • Google Vertex AI59.9%
  • Anthropic24.5%
  • OpenAI6.8%
  • Z.ai4.6%

Usage over time

19 Nov 2025 to 17 Nov 2026

Receipt · signedFetched 17 Nov 2026, 10:57 UTC
Explore the live dashboard

A sample account on live data, open to explore without a login

Every major AI provider and gateway, metered the same way

Start with the provider you use most. Connect it with a read-only key, an SDK, the MCP server, the CLI or the browser extension, and the first tokens land in seconds.

Coming soon: Cohere, xAI, Groq, Together AI, Replicate, Ollama.
One number, every surface

Count the tokens where they happen: the dashboard, your code, or your agents.

Dashboard, SDKs, MCP, CLI and API share one contract, so a token counted anywhere lands in the same report.

Dashboard

One view of tokens, cost and carbon across your providers, projects and coding agents, with CSV and JSON export.

v0.5.0

TypeScript SDK

A typed client for your usage, emissions, receipts and audit packs from Node.js or TypeScript.

npm install @crbonfree/sdk
v0.4.0

Python SDK

Sync and async clients for your usage, emissions and receipts from any Python service.

pip install crbonfree

MCP server

OAuth-secured remote tools, so your agent can read organization-scoped projects, connections, receipts and statements.

https://api.crbonfree.com/mcp
v0.2.4

CLI

Meter your local AI coding agents automatically, synced to the dashboard every hour.

npm install -g @crbonfree/cli && crbon install
v1.0.1

Browser extension

Estimates token usage in your AI chat sessions across ChatGPT, Gemini and the Claude apps, including Claude Code. Guest mode keeps everything on your machine, and signing in keeps the history.

Add to Chrome or Edge

AI Memory for Individuals and Teams

Search past coding agent sessions and AI chats, local by default with opt-in paid cloud features.

REST API

Twelve read endpoints for usage, emissions and receipts, with the OpenAPI spec published.

Numbers that hold up to questions

Signed retirement receipts

Each receipt lists the credit serial numbers on the American Carbon Registry or the BCarbon Registry.

Public verification

Anyone can check a receipt at verify.crbonlabs.com.

Monthly audit packs

A monthly zip with a manifest hash you can verify yourself.

Badge and certificate

Assets to back the claims you publish.

Receipts, audit packs and trust assets are on paid plans.

Case study · OpenRouterPublic traffic · Nov 2025 to Oct 2026

We asked the question about OpenRouter and metered a year of its traffic to answer it.

One gateway routes 787 trillion tokens a month, which works out to 173.7 kt of CO₂e, up 30x in a year, and none of it shows up in anyone’s Scope 3 disclosure. Your own curve has the same shape.

Tokens · monthly

787T

CO₂e · measured

173.7 kt

Growth · YoY

30x

Source: OpenRouter (openrouter.ai/rankings), as of 2026-11-17, fetched 17 nov 2026, 10:57 utc. Carbon computed with methodology v1.2 assuming a 75/25 input to output split.

Tokens routed per month

NovJanMarMayJulSep
26.6T in Nov 2025787T in Oct 2026

Sign up to measure your own usage.

However your team already runs AI, that is how you connect. Paste a read-only provider key, drop in an SDK, add the MCP server, install the CLI or switch on the browser extension, and the first measured tokens reach your dashboard within seconds.

Measure your AI for free

Free plan · first tokens in seconds

Before you connect anything

Is this just an estimate?

It is a measurement with a stated uncertainty. Every token is counted, converted with published factors that carry a version number, and reported with a plus or minus 28.3% band, the IPCC Tier 1 combination of 20% on activity and 20% on the factors, so you know how far to trust the figure.

Does CrbonFree see your prompts?

CrbonFree does not need your prompts to measure your AI, because measurement uses usage metadata only: the model, the token counts and the timestamps. AI Memory is different. It keeps the content of AI sessions so they can be searched later, and that content reaches Crbon only through the memory features that send it, such as memory sync or a Microsoft 365 Copilot connection.

What does the number include?

Three layers: the energy the model itself used, the data centre around it, and the hardware and training carbon behind it. Each layer is reported on its own, so nothing is blended in silently.

Questions

What people ask once they see the number.

These are the short answers. The longer ones, with the factors and the arithmetic, are on the methodology page.

Read the methodology
  • It depends on the model and the length of the exchange. Under methodology v1.2 a medium tier model spends about 0.3 joules per input token and 0.5 joules per output token on the accelerator, so a typical exchange of a few hundred tokens uses well under a watt hour of active energy before the facility around it is counted. The factors for every tier are published.

    See the per-tier factors
  • Joules become kilowatt hours, then CO₂e through the facility’s power usage effectiveness and the grid’s carbon intensity. Two further layers add host servers and cluster utilisation, then embodied hardware and model training, following Watershed’s open framework.

  • Yes. Data centres use water for cooling and the power plants behind them use more. The dashboard reports water alongside energy and carbon, derived from the measured energy, so the three numbers always describe the same usage.

  • They draw different boundaries. Some count only the accelerator, some the whole facility, some the hardware and training as well, and few say which. CrbonFree reports all three layers separately with the factor version on every figure, so two numbers can be compared on the same footing.

  • AI inference runs in someone else’s data centre, which makes it a purchased service under Scope 3. As disclosure rules tighten and data centres head toward roughly 12% of US electricity by 2028, auditors have started asking for the number, and a spend-based guess does not survive the question.

  • Yes. The methodology is public, with every factor, formula and source, and the sample account shows a year of real OpenRouter traffic metered through the same pipeline, with no login required.

    Open the sample account

Stop guessing the number your board is about to ask for.

Tracking usage, energy and carbon is free, and you will not need a card or a sales call.