Your AI bill arrives as one number. We meter it where it accrues.
Every agent call re-reads its whole conversation. Last month 93% of our own agent spend went to carrying context, and our detectors flagged about 30% of it as avoidable. Drop in your session logs and see your own numbers in about a minute.
Works with: Claude Code (CLI and Desktop) · OpenAI Codex CLI · GitHub Copilot · Cursor · Cline · Roo Code · Gemini CLI · Continue · any OpenTelemetry GenAI export. Claude Code is verified on real sessions; the rest are beta — send us a sample export and we'll verify yours.
Meter your sessions
Nothing is uploaded. The meter runs in this browser tab. It reads only usage fields (model, timestamps, token counts, cache split, tool names) and never your prompts, outputs or code. Tool inputs are reduced to a hash. You can disconnect from the internet after the page loads; the only file we send you is the price table.
Claude Code, Codex CLI, Gemini CLI, Cline, Roo Code, Continue: pick the log folder or drop the files. GitHub Copilot and Cursor: drop the usage export (.csv). Not sure where they live? See Where are my logs?
Prices: loading…
Claude Code is verified on real sessions. The other sources are beta: built from each tool's published source code and docs and tested on synthetic files. Using Copilot, Cursor, Codex, Cline or Gemini CLI? Send us a sample export (usage columns only) and we'll verify your format.
Your numbers
By source
Each source records different fields, so each runs only the detectors its data supports. A detector that can't run is listed as not applicable, never as a clean zero. Billing exports (Copilot, Cursor) are per-request account data: they show requests and the tool's own dollar figure, not per-session waste.
| source | files | sessions | context | output | $ list-eq. | $ tool-reported | requests | detectors run | not applicable |
|---|
Waste detectors
Sessions
| session | source | start | calls | context | output | $ list-eq. | $ tool-reported | carry | detectors |
|---|
Dollars are API list-price equivalents from our dated price table (). A flat subscription pays something else. A model not in the table is shown as tokens, never $0. "Tool-reported" is the figure Cline, Roo, Copilot or Cursor wrote themselves; we show it beside our list price and never add the two together.
Where are my logs?
Pick the folder or the files below. JSON files are read only by these known names, so a folder pick never opens a file that holds conversation content (for example Cline's api_conversation_history.json is skipped).
| tool | macOS | Windows | Linux |
|---|
Claude Desktop's Code tab runs the same engine as the CLI, so its local sessions land in the same ~/.claude/projects folder. Cloud and SSH sessions keep their logs on the machine that ran them. On macOS, press Cmd+Shift+. in the file picker to show hidden folders such as .claude and .codex.
The five detectors
- Context carry: output under 0.5% of context read across 50 or more calls.
- Context growth: context per call doubles across a session and passes 200K.
- Cache rebuild: a call rewrites at least half of a 50K+ context because the cache expired during an idle gap, so you pay the write premium again.
- Polling burn: four or more identical tool calls in a row, each re-reading the whole context.
- Fork drag: a subagent starts with its parent's history instead of a short brief.
Each detector ships only after it has fired on our own sessions. We killed a sixth, low-output polling, when it flagged 62 of our sessions for doing normal work.
Pricing
Pro
- Live: continuous metering, up to 25 seats, pushed with finops-push.js (usage fields only, never prompts)
- Live: weekly detector digest, dated snapshots and diffs
- Today: the same key runs our threat feed and Edge Shield
Enterprise
- Live: unlimited seats, team and feature chargeback
- Live: invoice reconciliation, token-theft detection (flags with reasons; never auto-revokes)
- Live, opt-in: AI-fluency insights (team-level, never a stack rank)
- Want it shaped around your team? Talk to us.
How it pays: a 10-seat team spending $600 per seat each month that captures half of a 30% waste finding saves about $900 a month against $99. For light users it may not pay, and the free meter will show you that before you spend anything. Session metering itself is not new; Claude Code exports usage telemetry. What we add is named waste detection on top, priced so a small shop can run it.