A reckoning for the environmental footprint of your AI usage.

Your organization is going to keep using AI, and it also wants to take AI's environmental cost seriously. Until now the choices were to ignore that cost or to overclaim about addressing it. R3CKON is the third option: an honest estimate of your usage-phase footprint with stated error bars, environmental certificates retired in response every month, and a record that says exactly what you can claim.

Free tier: estimates for up to 10M billable tokens a month. Email address only, no card. No API? Estimate from a few questions: for yourself or your whole organization.

Free accounts are open; paid tiers are not yet on sale, and the demo console below runs on demonstration data. Footprint figures are modeled estimates with stated 90 percent ranges, not measurements. Claims are activity claims: certificates record the restoration, reduction, and renewable generation you funded, identified by project and vintage. They never assert a neutrality status, and the certificate says exactly what statements it supports.

Metering to retirement, end to end

01

Meter

We count your AI usage from metadata alone: which models, how many tokens, which regions. Your prompts and outputs never reach us; anything content-shaped is rejected at the API boundary.

02

Estimate

No one outside the providers can know AI's environmental cost exactly, so we estimate it honestly: a public, versioned methodology turns your usage into water, carbon, and energy figures, each shown with its full range of uncertainty.

03

Retire

Every month, real money goes to work against the conservative end of that estimate (P90): third-party-issued certificates for water restoration, carbon reduction, and renewable energy are retired for your benefit, whole units only.

04

Certify

You receive a certificate naming exactly what was funded and retired (projects, vintages, supplier references), the estimate that sized it, and the precise statements you can publish on the strength of it.

The estimate and the retirement are separate assertions. Retirements are exact: registry-denominated instruments, retired on your behalf, documented with supplier references. Estimates are modeled, and their uncertainty is quantified and shown. The certificate keeps the two apart so the record your auditors see is the strong one.

The two-sided ledger

On the left, our best estimate of what your AI usage consumed, stated with its uncertainty because no one can know it exactly. On the right, the projects your subscription funded in response. Two columns, related but never equated, and never blended into one score.

LEFT SIDE OF THE LEDGER

Your usage, estimated

Providers do not disclose enough for anyone to know this exactly. This column is our best estimate of it, built from your token counts, with the error bars kept visible.

Fresh input tokensfull input coefficient
Cached input tokens≈10% of the compute; discounted in the estimate and toward your cap
Output tokensthe expensive direction (decode)
Reasoning tokensa subset of output, never double-counted; if a provider hides them, a conservative multiplier applies
feeds one estimate

Usage-phase footprint estimate

solid to P50 · lighter to P90, the accrual basis · whiskers span the 90% range

A ledger, not an equation. The right side responds to the left. It never equals it, and no claim we support says it does.

RIGHT SIDE OF THE LEDGER

Credits retired in response

Real money into projects that restore water, reduce carbon, and add renewable generation. The work happens where those projects are, which may be far from your compute, and the certificate says where.

WaterWRCs · 1,000 gallons each
Carboncredits · 1 tCO2e each
EnergyRECs · 1 MWh each

Whole registry instruments, identified by project and vintage on every certificate. Categories never mix: no combined score, no cross-category totals, and RECs are never presented as carbon credits.

Read it the way an auditor would. The left side is a modeled estimate with a stated range; the right side is exact instruments retired against that estimate's conservative end (P90). Restoring water in one watershed does not un-consume water in another, and a retired credit does not make a footprint disappear. What the ledger records, and what you may say, is the activity: estimated impact on one side, documented response on the other.
Metadata only, always. The platform ingests token counts, model names, and regions. It never proxies your AI traffic and never receives prompt or completion content; content-shaped payloads are rejected at the API boundary and enforced by tests.

Wired in however you run AI

The ledger needs token metadata and nothing else. Every path in is designed around one rule: we never accept a credential that could change anything in your account.

Pick your path

  • ·In your own code: npm install @r3ckon/sdk, wrap your client once. Exact cache and reasoning counts, and a failed report can never break your AI calls. Covers OpenAI-compatible gateways such as OpenRouter.
  • ·Already emitting OpenTelemetry? Point an exporter at our OTLP endpoint. No code changes, no credentials held.
  • ·Usage lives with your provider? A dependency-free sync script runs on your machine on a schedule; your provider key never leaves it.
  • ·No API at all? Answer a few questions and get an estimate with an honestly wide range, for yourself or org-wide.

Full trade-offs on Connecting usage.

Bring your coding agent

Most integrations are wired in by a coding agent now, so the documentation is written for one. Paste one instruction block into Claude Code or Cursor and it reads the docs, picks the path, and verifies its own work. Install the skill once and every future session already knows how.

Once connected, agents can read the live ledger over MCP: estimates with their ranges, retirement records, and the exact approved claim sentences, so the sustainability copy your agent drafts is the language the record actually supports. Read-only by design; no tool can push usage or change your account.

What you get, and what we refuse to sell

What you get

  • ·An honest monthly estimate of your AI usage's water, carbon, and energy cost, with the uncertainty stated instead of hidden.
  • ·A fixed retirement basket every month, whether you used your allowance or not, sized above the estimated footprint of your entire tier allowance. Your retirements meet or exceed the conservative end of your estimate by construction; the surplus is the point.
  • ·A certificate naming projects and vintages, plus the exact statements you can publish and stand behind under scrutiny.
  • ·A publishable activity record you can embed on your own site: confirmed retirements and estimates with their ranges, served from the ledger rather than pasted, so what your visitors see is what the record supports.
  • ·A price that reads as a small line item next to your AI bill, not a second one.

What this is not, on purpose

  • ·Not a neutrality badge. The footprint is an estimate with a wide stated range; declaring it settled would be fiction. The record shows activity instead: estimated cost on one side, funded response on the other.
  • ·Not a measurement. Providers do not disclose per-token energy or water. Anyone selling precise figures is guessing without error bars; we estimate conservatively, in the open.
  • ·Not always local. Verified projects live where they live: restoring one watershed does not un-consume water in another, so certificates state the geography rather than implying it away.

The longer answer, and where we intend to take it: our mission and vision.

Next: what we are trying to do, and what stands in the way

See the mission