Creekmark

Windows desktop app · free · runs entirely on your own machine

Every AI prompt burns power and water. This shows you how much — then helps you pay it forward.

Creekmark is a small app that lives in your Windows tray. It quietly counts the tokens your AI tools actually use, turns that into an honest estimate of energy, carbon and water, and lets you put real money against that impact — before you run it up, not as a guilt trip afterwards.

No account. No cloud. Nothing leaves your computer.

The Creekmark panel: today's usage of 46.2 million tokens shown as 0.21 kWh with a range of 0.04 to 2.77 kWh, a note that the high estimate is thirteen times the mid, carbon and water figures each with their own range, a 30-day sparkline, a funded balance of $24.41 covering about 44 months, and a Contribute button.
The panel, opened from the tray icon. Real output — the range is always shown, never a lone number.

It does three things

  1. 1

    It counts

    Every call your AI tools make reports how many tokens it used. The app reads those numbers off your own machine and logs them. This part is exact — it is a count, not a guess.

  2. 2

    It estimates — and shows you the doubt

    Tokens become watt-hours, grams of CO₂e and litres of water. Nobody publishes the real figures, so this part is genuinely uncertain — and the app always shows a range instead of pretending otherwise. On real usage the top of the range is about ten times the middle.

  3. 3

    You fund it forward

    It works out what your usage would cost to put right, and lets you buy a balance ahead of time that draws down as you work — so the impact is covered before it happens rather than owed afterwards.

A small always-on-top window showing today's energy figure and its range.
A tiny always-on-top window, if you want the number in the corner of your eye all day.
The methodology window, listing every conversion factor with its basis and confidence rating.
Every number has a “why”. One click from the panel, always reachable, never buried.

Pay it forward, not off

The usual model is to run up an impact and buy an offset later, if you remember. This one is built the other way round.

Buy a balance, let it draw down

A day of AI work might owe a few cents of remediation. No supplier will transact that, so waiting until it is “worth doing” means the impact happens with nothing funded against it at all. The app's answer is to fund ahead: buy a balance now, and watch it draw down as you use it. It tells you how many months of runway you have left and prompts you before it runs out.

Carbon and water are kept separate

They are not interchangeable, and a carbon contribution never reduces the water figure. Carbon is priced for durable removal, not avoidance — avoidance credits trade under $5 a tonne, and quietly defaulting there would make the suggested amount look reassuringly small while buying the thing that matters least.

It will never tell you that you are “carbon neutral”

That claim, and four others like it, are on a forbidden list the code checks every payload against — it throws if one ever appears, and a test proves the guard fires. The framing is fixed at “contributions logged against estimated impact”. Uncertainty this wide must not be laundered into a clean claim.

And the honest catch, stated in the app itself: replenishment — in the formal sense — means returning water to the catchment it came from, and your AI provider does not disclose which datacentre served your request. So you cannot call a water contribution replenishment of your usage, and this app will never say that you have. That is a limit on the claim, not a reason to skip it. A restored river is restored regardless of which one drew your inference. Fund it because it is worth doing, and let the claim stay what it actually is: money contributed, not water put back. The app says exactly this on the screen where you would spend the money.
The contributions window: a funded balance of $24.41 out of $25.00 contributed, 44 months of runway, state funded; suggested contributions of $0.59 for carbon and $0.03 for water, each shown with its full range; and Fund Ahead options of $20, $20 and $25.
The contributions screen. Funded balance, runway, and fund-ahead options — with the full range beside every figure.

Try the engine, right here

This is not a mock-up or a simplified copy for the website. It is the application's own conversion code, running in your browser against its own factor file. It re-checks itself on load and refuses to show any number at all if it disagrees with the app.

Start from

Estimated energy
mid · the high estimate is the mid
Each estimate on its own logarithmic scale — linear would pin low and mid against the left edge and make the estimate look far tighter than it is.
Where the energy goes — at the mid estimate. Shares differ at the bounds, because the terms do not share a divisor.

Non-text workloads take a different path entirely

The token chain does not describe speech-to-text, text-to-speech or image generation — different machinery, different physics. Running them through a text-model factor would produce a confidently wrong number, so each is estimated in its own unit, and every one of those factors is a labelled guess.

Estimates only, with wide error bars. Factor set , config .

Where the money could go

This is not a list of recommendations. These are candidates — somewhere to start reading, not somewhere to send money on our say-so. Creekmark itself will not recommend any of them to you until you have opened the link, read up, and marked it reviewed in the app. That gate is the whole point, and a website that quietly bypassed it would be worse than having no list at all.

So every entry below leads with what is wrong with it, at the same weight as the case for it. If you find that off-putting, that is the honest reaction — this is a market with real integrity problems, and the differences between these options matter more than the amounts involved.

Before you send anything anywhere. Creekmark cannot tell you which electricity grid served your prompts or which catchment the water came from — no provider discloses it. That is not a gap in this list; it is a limit on what any contribution against these figures can honestly claim. Read the water entries in particular with that in mind.

Candidate list version . Seeded from each organisation's public description and not checked against their current programmes, pricing, registry status or jurisdiction. Links open the organisation's own site so you can verify all of that yourself.

Get the badge for your website

The badge says one thing: this organisation funds the remediation of its AI usage, and that funding is backed by evidence. It carries no figures of any kind — no tokens, no energy, no carbon, no water, and no money. How much AI you use is nobody's business, and a published balance would imply it anyway. A trust mark should assert a fact, not publish a ledger.

Everything below is generated in your browser and is completely self-contained — no script, no tracking, no call back to anyone. Embedding it does not add a tracker to your site, and it does not count your visitors.

Also show
Your own standing

Let the app do this. Open Creekmark and choose Website badge from the tray menu — it reads your ledger directly and gives you the finished code to copy. Nothing to run, nothing to paste, and nothing leaves your machine.

This page cannot read your ledger by itself, and that is deliberate. The app listens only on 127.0.0.1; if a public website could reach it, so could every other website you visit. The badge above is a sample until you supply real figures.

Don't have the app open? Paste your ledger instead

Preview

“Verified” means there is a receipt. Every contribution logged in the app requires an evidence link — on an honour system that link is the only enforcement there is, so it is the price of entry rather than something to add later. Money without one still counts toward your balance inside the app, but the badge will not present it to the public as verified.
The badge is a snapshot, and it says so. It carries the date you generated it, because a static image on your website cannot know what you spent this morning. Regenerate it whenever you want it current — that is the honest price of a badge that phones nobody and tracks none of your visitors.

How the estimate is built, and where it is weakest

The obvious approach is two separate figures — one for what you send, one for what the model writes back. That makes it far too easy to encode a ratio the published measurements contradict. Generating text costs somewhere between 17 and 200 times more energy per token than reading it, and which end of that range applies depends on how busy the provider's servers are, which you cannot see from outside.

So the app anchors on the generation figure and derives the reading figure from it. It also carries a term most estimates leave out entirely: the longer the conversation, the more energy each new word costs, because the model re-reads the whole context to produce every one. On long agent sessions that term is the single biggest line item.

Three things the code refuses to do

Work out the cache figure from the price
Cached input is billed at a tenth of the normal rate. That is a commercial decision about margin, with no established relationship to actual energy. The note in the config says so, so nobody later “confirms” the number against a pricing page.
Quietly carry on with a broken config
A malformed factor file is rejected outright at load. A plausible wrong number on screen is the worse failure.
Average out its confidence
Confidence is the lowest of everything that fed the number, never the average — averaging would let three solid inputs bury one that has no published basis at all.

Ranked weakest links — shipped in the app, not buried in a whitepaper

What it cannot tell you

On the record, on the front page. A tool about not overstating things does not get to overstate itself.

  • The long-conversation figure is the weakest number in the app

    That energy rises with conversation length is well established. How steeply is not. No provider publishes it, the figure used is order-of-magnitude plausible and nothing more, and it is flagged as such on every result and listed first in the app's own provenance view.

  • Nobody publishes per-token energy

    The model factors are extrapolated from the handful of published per-prompt figures that exist. Figures for non-Anthropic models are labelled outright guesses. They are included because leaving that usage out would understate your total in a known direction — worse than a labelled estimate.

  • It cannot tell where your request was served

    Providers don't disclose it, so the carbon figure uses a broad regional average rather than the grid that actually ran your prompt, and the water figure cannot name the catchment it came from. That limits what a water contribution may be called — not whether it is worth making.

  • Training is not in these numbers

    This covers running the calls you made. Building the models in the first place, your own computer, and moving the data around are all outside it.

Get the widget

There's a working Windows installer. It isn't code-signed yet, so Windows would throw a scary warning at anyone who downloaded it from an open link — so it goes out by email instead, with install notes. Tell me where you'd run it and I'll send it over.

Worth knowing first

  • Windows. Built and tested on Windows 11. The engine underneath isn't Windows-specific, but the tray app hasn't been exercised anywhere else.
  • It runs entirely on your machine. No account, no sign-up, no data sent anywhere. If this site vanished tomorrow your copy would carry on working.
  • It is free, and there is nothing to upsell you. The only money involved is what you choose to contribute, and that goes to the recipient, not to me.
  • It will not tell you that you are carbon neutral. That claim is unavailable by construction.

More from BrambleRidge Home Design Group