how tokens are counted

Every token your Claude Code sessions actually use: input + output + cached context (cache reads and writes). It's the same total Claude Code shows in its live counter, and the same tokens Anthropic bills — including usage from subagents and background tasks.

Your machine reports usage through each tool's official OpenTelemetry metrics — a first-party, supported telemetry channel, not a fragile hack. We receive token counts and timestamps, nothing else.

We never see your prompts, responses, code, or any conversation content. It is numbers only. Tracking is opt-in and you can disconnect any machine at any time.

In agentic coding, most of the work is re-reading context, so cache reads make up the bulk of real usage. Excluding them would show a tiny, misleading sliver of what you actually did. We count the true, billable total — so a big day looks like a big day.

Claude Code — terminal and desktop app — via its official OpenTelemetry metrics (input + output + cache reads and writes, including subagents). OpenAI Codex CLI via its official telemetry too: input, output, cached, reasoning, and tool tokens — the same billable total Codex itself reports.

We only track tools that expose usage through a stable, official mechanism, so the count is accurate and consistent on every machine you connect. We don't track plain claude.ai or ChatGPT chat because there's no reliable way to measure it — and we won't show a number we can't stand behind.

Because the count comes straight from Claude Code's own telemetry, it reflects real work. We also cap and sanity-check ingestion so a fabricated payload can't rocket someone up the board.