Reads the local logs of Claude Code, Codex, ZCode and 19 agents total — read-only, incrementally indexed into SQLite, rolled up into a fixed set of views chosen for you. Local-first, reconcilable.
Windows · macOS · Linux · AGPL-3.0
JSONL here, session snapshots there, SQLite elsewhere. Mixed together, you honestly cannot say what today cost.
Read-only access to 19 sources, auto-detected with zero config. No external processes spawned.
19 formats → 1 ledger ccusage cold-starts take tens of seconds — a coffee-break to check today’s spend.
Files are tracked by mtime/size; only changes get re-scanned, landing incrementally in local SQLite. Opens with numbers ready.
full rescan → incremental An estimate is an estimate — folding unpriced models into totals as zero is the real problem.
Missing prices show “unknown” — never zeroed into totals. costUSD from logs is used as-is.
unknown ≠ 0 Measured on real data (local Codex: 19 files / 2,931 records) — cent-exact against ccusage over the full period.
vs npx ccusage@20.0.29 · day × model reconciliation
All five token components match exactly
Reconciliation tests parse only the golden JSON committed to the repo — repeatable offline.
A fixed information architecture — no plugins, no layout customization. Overview, sessions, agents, models, projects, time, cost, data & pricing, settings, with time-range and four-dimension filters running through every view.
Sessions: agent, dominant model, project, start/end, tokens and cost — row by row, with search, filters, sorting and paging. Unpriced sessions show “unknown”, honestly.
Cost analysis: daily cost trend, model composition and period-over-period deltas, plus the three-line methodology — costUSD as-is, unknown never zeroed, reconcilable.
Models: provider, associated agents, token and cost shares in one table. Unpriced models carry an “unknown” mark — honesty over cosmetics.
Time analysis: hourly distribution (sources without minute-level precision count as unknown — no fabricated shape), weekday spread, activity calendar and time KPIs.
Projects: share, sessions, averages — click through to that project’s sessions.
Agents: top KPIs and a ranked table — click through to that agent’s sessions.
Your overrides → models.dev refresh (background, never blocks the UI) → bundled price snapshot. Source status, record counts and the model catalog on one page.
Plus 09 Settings — theme, language, sources and price overrides all in plain sight.
No need to open the main window: today’s tokens, estimated cost, cache-hit rate and live rate at a glance; abnormal spend gets a badge.
All read-only access (streamed JSONL/JSON parsing or SQLite read-only connections), auto-detected with zero config, overridable via environment variables.
Where data lands
%LOCALAPPDATA%\Tokenmetry\ user data directory SQLite stores only aggregated facts and settings — never prompts, code or conversation content.
To be honest: the only two network touchpoints are the models.dev background refresh and the in-app update check.
AGPL-3.0-or-later (brand name and logo not covered). The source is the product — clone and run.
Platform packages are built by tag-triggered CI; before the first public release, please run from source. Prerequisites: .NET 10 SDK and local logs from any supported agent.
Run from source (primary today)
git clone https://github.com/cholf5/tokenmetry.git
cd tokenmetry
dotnet run --project src/Tokenmetry.App