toktrack

English | 한국어

⚠️ Did you know? Claude Code deletes your session data after 30 days by default. Once deleted, your token usage and cost history are gone forever — unless you preserve them.

The token & cost tracker that never loses your history. Most tools re-read your CLI's session files on every run — so when Claude Code deletes them after 30 days, your cost history goes with them. toktrack keeps a persistent cache, so your history survives.

Track usage across all your AI coding CLIs — Claude Code, GitHub Copilot CLI, Codex CLI, Gemini CLI, Qwen Code, OpenCode, PI Agent, Antigravity, and Grok CLI — in one dashboard. Built in Rust, so it stays fast even on huge histories (simd-json + rayon).

• Data Preservation — Persistent cache keeps your cost history even after a CLI deletes its own session files
• Multi-CLI Support — Claude Code, GitHub Copilot CLI, Codex CLI, Gemini CLI, Qwen Code, OpenCode, PI Agent, Antigravity, Grok CLI in one place
• TUI Dashboard — 5 tabs (Overview, Stats, Models, Projects, Audit) with daily/weekly/monthly views
• Per-Project Breakdown — the Projects tab shows token & cost usage per project (the session working directory), for CLIs that record one. Drill into any project for its day-by-day, per-model breakdown. Usage from CLIs that don't record a project is grouped under (no project). Project details are cached, so they survive past the CLI's 30-day deletion
• CLI Commands — daily, weekly, monthly, stats with JSON output support
• Usage Reports — Shareable text & SVG receipts via toktrack report
• Fast on huge histories — simd-json + rayon parallel parsing (~3 GiB/s); cached runs in ~0.04s

No Rust toolchain required. Downloads the correct binary for your platform automatically.

npx toktrack
# or
bunx toktrack

brew tap mag123c/toktrack
brew install toktrack

cargo install --git https://github.com/mag123c/toktrack

Download from GitHub Releases.

# Launch TUI dashboard
npx toktrack

# Get today's cost in JSON
npx toktrack daily --json

# Monthly summary
npx toktrack monthly --json

toktrack

# Open TUI at specific tab
toktrack daily     # Overview (daily view)
toktrack weekly    # Overview (weekly view)
toktrack monthly   # Overview (monthly view)
toktrack stats     # Stats tab

# JSON output (for scripting)
toktrack daily --json
toktrack weekly --json
toktrack monthly --json
toktrack stats --json

# Usage report (shareable receipt)
toktrack report              # Last 7 days (text)
toktrack report --month      # Last 30 days
toktrack report --days 14    # Last N days
toktrack report --svg        # Text + SVG file

# Data-preservation audit (live on disk vs cache-only)
toktrack audit               # Per-source coverage report
toktrack audit --json        # Machine-readable

Show today's AI spend right in your prompt, updated after every command:

~/projects/api  main  $12.40
❯

A few lines of zsh/bash/fish — see Shell integration.

toktrack can include Codex session logs from SSH servers by syncing remote JSONL files into a local snapshot and parsing them with the existing Codex parser. It uses your existing ssh/rsync setup and never stores SSH passwords or keys.

Create ~/.toktrack/config.toml:

default_remotes = ["devbox"]

[[remotes]]
name = "devbox"
target = "ubuntu@devbox"

[remotes.paths]
codex = "~/.codex/sessions"

[[remotes]]
name = "prod"
target = "prod-alias"

[remotes.paths]
codex = "/home/codex/.codex/sessions"

You can also manage the same config from the CLI:

toktrack remote add devbox ubuntu@devbox --default
toktrack remote add prod prod-alias --codex-path /home/codex/.codex/sessions
toktrack remote default add prod
toktrack remote default remove devbox
toktrack remote remove prod
toktrack remote list

By default, commands read local data plus default_remotes:

toktrack report
toktrack daily --json

Override the configured defaults when needed:

toktrack --remote prod report      # local + default_remotes + prod
toktrack --all-remotes stats       # local + every configured remote
toktrack --local-only report       # local only

Remote Codex sources appear as codex@<remote-name> and use separate cache files such as ~/.toktrack/cache/codex@devbox_daily.json. If a remote sync fails, toktrack warns and continues with local data plus any existing snapshot/cache for that remote.

Costs from sources that don't record their own price (Gemini, Qwen, Codex, Antigravity, GitHub Copilot, and modern Claude logs) are computed from LiteLLM API list-price estimates and shown with a ~ marker (estimated). For GitHub Copilot CLI, the displayed value is the API-equivalent cost rather than your actual subscription/credit spend. Sources that do record their own cost (Grok CLI, OpenCode, PI Agent, and older Claude logs) have it used verbatim rather than estimated. The ~ marker is per source, not per entry: if any one entry in a source lacks an upstream cost, the whole source is marked estimated. A Grok turn that reports no cost of its own falls back to the pricing table, which has no bare grok-4.6 key, so it contributes $0 to the totals. Pricing works offline via a bundled snapshot when the network is unavailable. 1h ephemeral cache writes are billed at their own higher rate, and Claude Code requests made in fast mode (/fast) are billed at the provider's fast multiplier. A model LiteLLM has no price for shows ? instead of a cost, so a $0 row is never mistaken for free usage.

Codex Fast costs follow the thread_settings_applied.thread_settings.service_tier snapshots a session recorded: priority and legacy fast enable Fast pricing, other or missing tiers use Standard. That event is the only place Codex writes the tier, and many sessions contain none of them, so usage with no tier on file stays Standard rather than inheriting today's local config. Expect Fast pricing to show up only for the sessions that recorded it, not for every session you ran in Fast mode. Known model multipliers match ccusage's model table (GPT-5.5: 2.5×; GPT-5.3-Codex, GPT-5.4, GPT-5.6/Sol/Terra/Luna, and GPT-6 Astra: 2×), and a model with no published Fast rate stays at Standard. These are dollar cost estimates, not ChatGPT credit balances. Upgrading recalculates dates with available logs while preserving cached history whose logs are no longer available.

Each source's data directory can be overridden (matching the upstream CLI's own variable):

export CLAUDE_CONFIG_DIR="/path/to/.claude"

Custom pricing (including web_search / web_fetch per-request rates) can be set in ~/.toktrack/pricing.toml.

Measured on Apple Silicon with 2,000+ JSONL files (3.4 GB). The persistent cache means only the current day is recomputed on each run — past days are read straight from cache, so daily use stays instant no matter how large your history grows.

Why it's fast: SIMD JSON parsing (simd-json) + parallel processing (rayon) = ~3 GiB/s throughput on the cold path.

The Problem: You've been using Claude Code for 3 months, spending hundreds of dollars. One day you want to check your total spending — but Claude Code already deleted your session files from 2 months ago. That cost data is gone forever.

toktrack solves this. It caches daily cost summaries independently, so your usage history survives even after the CLI deletes the original files.

~/.toktrack/
├── cache/
│   ├── claude-code_daily.json   # Daily cost summaries
│   ├── codex_daily.json
│   ├── codex@devbox_daily.json
│   ├── gemini_daily.json
│   ├── opencode_daily.json
│   └── pi-agent_daily.json
├── remotes/
│   └── devbox/codex/sessions/   # Synced remote Codex JSONL snapshots
├── config.toml                  # Optional remote source configuration
└── pricing.json                 # LiteLLM pricing (1h TTL)

Past dates in each *_daily.json are immutable — once a day is summarized, the cached result is never modified. Only the current day is recomputed on each run. This means even if Claude Code deletes session files after 30 days, your cost history remains intact in the cache.

toktrack audit shows, per source and per day, whether the data is still live (raw session file on disk), cache-only (the CLI deleted the raw file — only toktrack still has it), or missing (no data — an unused day, or one lost before toktrack first saw it; never claimed as a loss).

$ toktrack audit

Data preservation audit (2026-06-08)

  claude-code   2025-12-22 → 2026-06-08
    live (raw on disk):         79 days
    preserved (CLI deleted):    83 days
    no data (unused or lost):    7 days
  ...

  ► 100 days of cost history preserved that your CLIs already deleted.

Use toktrack audit --json for the full per-day breakdown, or open the Audit tab in the TUI (press 4) for a visual coverage map.

// ~/.claude/settings.json
{
  "cleanupPeriodDays": 9999999999
}

rm -rf ~/.toktrack/cache/

The next run will rebuild the cache from available session data.

Cold path (first run): Full glob scan → parallel SIMD parsing → build cache → aggregate.

Warm path (cached): Load cached summaries → parse only recent files (yesterday midnight mtime filter) → merge → aggregate.

Deep Dive: I Rewrote a Node.js CLI in Rust — It Went from 43s to 1s | 한국어

make check    # fmt + clippy + test (pre-commit)
cargo test    # Run tests
cargo bench   # Benchmarks

Now tracking 8 CLIs (Claude Code, Copilot, Codex, Gemini, Qwen Code, OpenCode, PI Agent, Antigravity) — see Supported AI CLIs. Planned: live/burn-rate monitoring, MCP server / statusline, more CLIs (Goose, Amp, Kimi).

Issues and PRs welcome!

make check  # Run before PR

MIT