Token and cost statistics for local Claude Code and Codex (OpenAI) sessions, priced at API list rates. Ships as a terminal TUI, a live browser dashboard, and native apps for macOS and Windows.

The engine is a single dependency-free Python file. The rest is not Python: the dashboard is plain HTML/CSS/JS (gui/), the macOS app is Swift + WKWebView (macos/), the Windows window uses WebView2.

Edit the dashboard in gui/, run python3 tools/embed.py, commit both.

./ccost              # interactive TUI (live: numbers grow as agents work)
./ccost gui          # live dashboard in the browser
./ccost app          # same dashboard in a native window (Win/Linux/mac)
./ccost report       # full text report (also the pipe/non-tty fallback)
./ccost json         # aggregates as JSON, for scripts

Prebuilt binaries (no Python needed) are in releases: ccost-macos.dmg (app), macOS universal2 / Linux x86_64 / Linux arm64 / Windows CLI binaries, and ccost-windows-x64.exe — a windowed app, double-click to run.

Local server on 127.0.0.1, everything inline (no CDNs, works offline). Live updates every few seconds via the incremental parser. Sections: cost counter with live "+$" ticks, stat cards (today + last hour, month forecast, cache savings, average day), daily bar chart with day drill-down (click a bar, ←/→ to page days, Esc to reset), week × hours heatmap, models, top chats, tools, hours of day, a GitHub-style year calendar, and all-time records (priciest hour/day/chat, longest chat, best streak…).

Settings (gear icon):

• sources — Claude Code and Codex toggles, with paths and counts;
• refresh — polling interval (2/5/10 s);
• default period, update checks;
• menu bar counter (mac) — pick what it shows: $ today, $/hour, $ month, messages, tokens (any combination);
• prices — edit any model's rates in place; the whole history is repriced instantly.

Config, the parsed-state snapshot and the webview profile all live in one folder, ~/.ccost/ (override with CCOST_HOME). Move it to a new machine and the whole history comes with it — no dependency on ~/.claude or ~/.codex. Older installs are migrated into it automatically.

Data paths can be overridden via claude_root/codex_root in the config or the CCOST_ROOT/CCOST_CODEX_ROOT environment variables.

Everything ccost keeps — config, the parsed-state snapshot, the window profile — lives in one folder, ~/.ccost/ (override with CCOST_HOME). Move that folder to a new machine and your whole history comes with it, with no dependency on ~/.claude or ~/.codex.

ccost.app is a native shell: full-height window (the dashboard header is the titlebar, drag it to move, double-click to zoom), Dock icon, Cmd+Q/W/R, and an optional menu bar counter (the window can be closed while the counter keeps running). The counter's menu configures it in place: pick the metrics it shows ($ today, $/hour, $ month, messages, tokens) and toggle Launch at login — started that way, the app stays in the menu bar and does not open the window. The bundled engine is a PyInstaller binary — no system Python required.

CCOST_ENGINE=/path/to/binary sh macos/build.sh   # or let it run pyinstaller
cp -R macos/dist/ccost.app /Applications/

ccost-windows-x64.exe is a real app: double-click opens a frameless window (the dashboard draws its own titlebar and — ▢ ✕ buttons, WebView2). For terminal use there is ccost-windows-cli.exe (TUI glyphs default to ASCII; set CCOST_UTF8=1 in Windows Terminal for unicode bars). If Claude Code lives in WSL, point the exe at it: set CCOST_ROOT=\\wsl.localhost\Ubuntu\home\<user>\.claude\projects.

Tabs: Overview · Models · Days · Chats · Projects · Tools · Hours · Week. Mouse works: click tabs and rows, wheel scrolls, second click opens details.

• prices — API list rates ($/1M tokens) for Anthropic (PRICE) and OpenAI (PRICE_OPENAI, cached input priced separately); model ids match by longest prefix; unpriced models count as $0 and are listed explicitly;
• cache — Anthropic 5-minute writes ×1.25, 1-hour writes ×2 (split via usage.cache_creation), reads ×0.1 of the input rate;
• dedup — streaming writes several JSONL lines per message with growing usage; the field-wise maximum per message.id is kept (the first line undercounts output, and resumed/forked sessions would double-count);
• Codex — usage comes from token_count events in rollout files, model and project from turn_context/session_meta, resume duplicates dropped.

Note: these are API list prices — what the usage would cost, not what a subscription actually charges (same numbers /status shows).

The full JSONL parse (gigabytes) happens once: parsed state is snapshotted to ~/.ccost/state.pickle and later runs start in a fraction of a second, reading only newly appended bytes. Cold parses run across several processes. Prices are not part of the snapshot, so repricing never triggers a re-parse.

The palette adapts to the terminal (256-color, 8/16, or mono), the background stays native, and UTF-8 glyphs are enabled before initscr so bars do not degrade to ? even without LANG. NO_COLOR=1 disables colors, CCOST_ASCII=1 forces ASCII graphics. Chat titles are sanitized (ANSI codes stripped, wide glyphs replaced) so layouts never break.