Scope 3 · AI usage · Connector-native
Every AI call your team makes, counted and graded.
Your people are already using Claude every day, and none of it appears in your carbon accounting. WattScope attaches as a connector, attributes each model call to an organisation and a project, and reports the estimated footprint — with the strength of the evidence behind every figure stated on the figure itself.
Get started See how figures are graded →
No API key No vendor credential handed over Nothing for security to review
Your team, in Claude
Ordinary work, in the tool they already use. Nobody changes how they do their job.
WattScope, riding along
- Attributed to an org and a project
- Graded by where the count came from
- Converted to an estimated footprint
A figure you can defend
Dashboard, CSV export, or ask Claude directly. One number, and the grade that travels with it.
Every figure carries its source grade
Where a number came from decides what it can be used for. WattScope never averages across that boundary, and never quietly promotes a weaker figure to a stronger one — the grade travels with the row, from the moment it is recorded through to the export.
| Grade | What it means | Class |
|---|---|---|
| Vendor Admin API | Token counts pulled from the vendor's own usage API, reconcilable against the bill. The strongest provenance available. | Billing-grade |
| Transcript sync | Token counts read from the conversation transcript the vendor returned. Machine-recorded, but not reconciled against a bill. | Indicative |
| Client-measured (OTLP) | Token counts exported by the AI client's own instrumentation as the work happened. Machine-recorded, and trusted as far as the client is. | Indicative |
| Self-reported | Token counts supplied by a person or an agent calling record_usage. Unverified: nothing checks them against the vendor. | Indicative |
Only the top grade is reconciled against a vendor's own invoice. Everything else is indicative, however precisely it was recorded — a token count read from a machine is still not a token count anyone has billed.
How it works
1. Connect once. Add WattScope as a connector in Claude and sign in here. No API key, no credential handed over, nothing for your security team to review — the connector authenticates you, not your vendor account.
2. Usage is attributed as it happens. Calls are recorded against your organisation and, where the client reports it, the project they belong to. Exact per-call token counts arrive through your AI client's own telemetry export, so nothing has to be estimated that can be counted.
3. Read it back anywhere. Ask Claude directly, open the dashboard, or export CSV. Every surface shows the same graded split, because they all read it from the same place.
What this does not claim
These are estimates, not measurements. No provider publishes per-token energy figures for hosted models, so the energy behind each call is inferred from public analyses of whole-query energy use and is uncertain by a factor of several in either direction. WattScope reports the estimate, the assumptions behind it, and the uncertainty — it does not round any of them away.
The factor table is still a draft. The deployment you are
reading is running factor set 2026.08.18a-draft, whose own
status line reads DRAFT - REVIEW BEFORE ANY EXTERNAL OR COMPLIANCE USE. The same string is returned
to every client that calls server_info. Until that changes,
do not use these figures for regulatory reporting, public claims, or offset
retirement volumes without replacing the table with figures your
organisation can defend.
We would rather tell you this on the homepage than have you discover it in an audit. Water use is measured by almost nobody and is on the roadmap for the same reason — it ships when the model can support it, not when it would look good here.