🌪️ Toknado

A tornado of tokens. See exactly what your coding agents devour — Claude Code, Codex and GitHub Copilot Chat, one dashboard, 100% local.

Toknado reads the session logs that Claude Code (~/.claude) and Codex CLI (~/.codex) already write to your machine, and turns them into a fast local dashboard. No accounts, no uploads, no telemetry, no database. Close the tab and nothing remains but your logs.

npx toknado

That's it. Your dashboard opens at http://127.0.0.1:4141.

What you get

• 📊 One dashboard for all your agents — Claude Code, Codex CLI and GitHub Copilot Chat (VS Code), merged or filtered per source
• 🔬 Every model, every call — the All-models table lists each model with API-call counts, avg tokens/call and full usage breakdown
• 📅 Every range you'd ever ask for — today, 7 / 14 / 30 days, all time, or a custom date range
• 📈 Charts, not spreadsheets — daily stacked bars (by token type or source), model share donut, activity heatmap, busy-hours histogram
• 🤖 Per-model breakdown — which model ate how much, with share percentages
• 🧠 Modes visible — Codex reasoning efforts (medium / high / xhigh / max / ultra) and Claude response speeds (standard / fast)
• 📉 Real quota tracking — Codex logs its rate-limit used_percent; Toknado charts your actual weekly quota consumption over time
• % of everything — every day, session, model, and project shows its share of total usage
• 🧾 Top sessions leaderboard — which sessions and projects burned the most
• 💸 What-if API cost comparison — what your token mix would cost at public API list prices, across model tiers (clearly labeled hypothetical — subscriptions don't bill per token)
• 📡 Live pricing (opt-in) — --live-pricing fetches current rates for thousands of models from LiteLLM's public price DB, so even brand-new models get priced
• 📤 Export on demand — JSON, CSV, or Markdown. Nothing is saved unless you export it

Privacy model

| Question | Answer | |---|---| | Where does my data go? | Nowhere. The server binds to 127.0.0.1 only. The single exception: with the opt-in --live-pricing flag, Toknado makes one plain GET to LiteLLM's public price file — it carries nothing about you. | | What does Toknado store? | Nothing. Every view is recomputed from your logs in memory. | | What if I close it without exporting? | Nothing is saved — and nothing is lost. Reopen it and the same history is there, because your CLI logs are the data. | | Can it modify my logs? | No. Logs are opened read-only. |

How it works

• Claude Code writes one JSONL file per session under ~/.claude/projects/. Each assistant turn records exact API usage (input / output / cache read / cache write) per model. Toknado deduplicates entries that appear in multiple files (forked / resumed sessions).
• Codex CLI writes rollout JSONL files under ~/.codex/sessions/. Forked/subagent files copy the parent's entire history with rewritten timestamps — Toknado deduplicates on usage snapshots so each API call counts exactly once, reads model + reasoning effort from turn_context, and collects the rate_limits.used_percent samples for the quota chart.
• GitHub Copilot Chat (VS Code) stores per-session operation logs under the editor's workspaceStorage/*/chatSessions/. Toknado replays the op log to recover each request's promptTokens/completionTokens, model and timestamp (Code, Insiders and VSCodium are auto-detected).

Token counts come straight from the logs — they're exact. Dollar figures in the comparison panel are estimates from public list prices.

Usage

npx toknado [options]

  -p, --port <n>        port to serve on (default 4141)
  --no-open             don't auto-open the browser
  --claude-dir <path>   Claude Code projects dir (default ~/.claude/projects)
  --codex-dir <path>    Codex home dir (default ~/.codex)
  --copilot-dir <path>  VS Code workspaceStorage dir for Copilot Chat
  --pricing <file>      JSON file with per-Mtok price overrides
  --live-pricing        fetch current API list prices from LiteLLM's public DB
                        (the only network request Toknado can ever make; off by default)

Custom pricing

Models Toknado doesn't know show n/a in the cost panel. Provide your own rates (USD per million tokens, prefix-matched against model IDs):

{
  "gpt-5.6": { "input": 1.25, "output": 10 },
  "my-local-model": { "input": 0, "output": 0 }
}

npx toknado --pricing my-prices.json

Non-goals

• Not a spend tracker. Subscription plans don't bill per token; Toknado is tokens-first everywhere, and the single cost panel is explicitly hypothetical.
• Not a sync service. There is no cloud, and there never will be.

Requirements

Node.js ≥ 18. Zero dependencies.

Contributing

Contributions are very welcome — see CONTRIBUTING.md. Good first areas: new log formats (other agent CLIs), chart ideas, and pricing-table updates.

License

MIT