🔥 ai-burn

What did your AI actually burn?

The CO₂, water and money your AI coding tools quietly torched — in one command, fully local, nothing uploaded.

npx ai-burn

  🔥 ai-burn — what your AI burned (all time)

  9.82 kg CO₂e   ·   49.1 L of water   ·   $709.55 at API prices

  🚙 Daily Commuter
  ███░░░░░░░░░░░░░░░░░░░ 15% to 🚛 Diesel Hauler (20 kg)

  That's like…
  🚗 58 km in a petrol car
  🍔 3.9 cheeseburgers
  🎈 545 party balloons of pure CO₂
  🧊 295 cm² of Arctic sea ice, gone

  …and your machines drank
  ☕ 1.64k espresso shots, evaporated
  🥤 98 half-litre bottles
  🌊 0.000000000055% of Lake Geneva

  By tool
  ✳️ Claude Code ██████████░░    8 kg  81%  $664.00
  🔓 OpenCode    ██░░░░░░░░░░ 1.73 kg  18%   $40.75
  🌀 Codex CLI   ░░░░░░░░░░░░  56.1 g   1%    $2.79
  🖱️ Cursor      ░░░░░░░░░░░░  28.3 g   0%    $1.77 ≈
  ♊ Gemini CLI  ░░░░░░░░░░░░  4.31 g   0%    $0.24

  ⚡ 20 kWh   🔤 1.02B tokens   📅 23.9 g/day   🔥 peak day: 1.91 kg (Jun 6)
  🏆 top model: claude-opus-4-8 (43%)

  Rough estimates — run with --json for raw numbers, README for methodology.

Everything runs locally: it reads the session history your AI tools already keep on disk, counts your tokens, and converts them to energy, emissions, water and money. Nothing is uploaded anywhere — ai-burn never makes a network request. The output is information, plain and simple: what you do with it is up to you.

🇫🇷 Parle aussi français — --lang fr (auto-detected from your locale).

🌐 Website · 🔬 Methodology · 🔒 Privacy · ⭐ Star this repo

Supported tools

| Tool | Where it reads | Token counts | | --- | --- | --- | | ✳️ Claude Code | ~/.claude/projects/**/*.jsonl | exact (per API turn) | | ♊ Gemini CLI | ~/.gemini/tmp/*/chats/*.json | exact | | 🌀 Codex CLI | ~/.codex/sessions/**/*.jsonl | exact | | 🖱️ Cursor | …/Cursor/User/globalStorage/state.vscdb | exact tokens, model unknown (≈) | | 🔓 OpenCode | ~/.local/share/opencode/opencode.db | exact (+ its own cost figures) |

Cursor and OpenCode are SQLite databases, so those two need the sqlite3 CLI on your PATH (preinstalled on macOS; apt/dnf install sqlite3 elsewhere). Missing tools are silently skipped. Using something we don't read yet? PRs welcome — a source is ~60 lines in src/sources/.

Options

| Flag | Effect | | -------------- | --------------------------------------------------- | | --days <n> | Only count the last n days | | --tools <ids>| Subset of claude,gemini,codex,cursor,opencode | | --lang <code>| Report language: en, fr (default: your locale) | | --json | Machine-readable output (always English keys) | | --no-color | Plain text (NO_COLOR also respected) | | -y, --yes | Skip the confirmation prompt |

Privacy & safety

• Asks first. Before reading anything, ai-burn asks for a Y/n confirmation in your terminal (skip with --yes; non-interactive runs proceed, since there is no one to ask).
• Fully local. No network access at all — zero HTTP requests, no telemetry, no analytics, nothing phones home. The only runtime dependency is Node's standard library (plus the sqlite3 CLI for two of the sources).
• Never touches credentials. The tool databases it reads also contain auth tokens (e.g. OpenCode's account table). The queries here only read chat/usage rows — never auth tables, never auth.json, never key files.
• No content leaves the math. Your prompts and code are never parsed for meaning, stored, or displayed — only token counts, models and timestamps.

Methodology (a.k.a. how wrong is this?)

Vendors publish almost no per-token energy figures, so this is an estimate, triangulated from public LLM-inference studies (Epoch AI's GPT-4o estimate, Google's Gemini per-prompt disclosure, Mistral's Large 2 lifecycle report):

1. Tokens are read per API turn from each tool's local history and deduped across resumed/forked/checkpointed sessions.
2. Energy: each model gets a Wh-per-million-tokens rate by tier (frontier — Opus/Fable-class: 1200 Wh/M output; standard — Sonnet/GPT-5/ Gemini-Pro-class: 500; light — Haiku/Flash/mini-class: 100). Input tokens count at 1/10 of the output rate (prefill is much cheaper), cache reads at 1/100, cache writes like input. Datacenter overhead (PUE) of 1.2 applies.
3. Carbon: 0.4 kg CO₂e per kWh (world-average grid mix).
4. Water: 2.0 L per kWh — on-site evaporative cooling (~1.1 L/kWh, Google's fleet WUE) plus water consumed generating the electricity (~0.9 L/kWh). See Making AI Less Thirsty (Li et al., 2023).
5. Money: retail pay-as-you-go API list prices per model (src/pricing.js), with cache reads at ~0.1× and cache writes at ~1.25× the input price — i.e. what your usage would cost if you weren't on a subscription. OpenCode records its own per-message cost; we trust it when present.

Real numbers depend on hardware, batching, datacenter location, grid mix and the phase of the vendor's pricing moon — expect the truth to be within a factor of a few, in either direction. All constants live in src/co2.js and src/pricing.js; PRs with better-sourced numbers are welcome.

License

MIT — made by Théophile J-R.