🐷 Watthog

Your AI coding agents eat electricity. Meet the pig that counts it.

One command. No API keys, no sign-up, no data leaving your machine. Watthog reads the logs your coding agents already wrote and tells you the electricity, CO₂e and water they actually burned — every number inside an honest low–high band, never a single digit you're asked to trust.

npx watthog

watthog.vercel.app  ·  How it works  ·  What it reads  ·  Methodology

🐷 What you get

🐷 watthog · estimated electricity use of your LLMs
16,187 assistant messages since 2025-05-06

     \   ^..^_________     SUBSTATION SWINE  ·  Lv 27
   --+-- ( ●  ●       )    ▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░ 74% → Lv 28
     /   (            )~   power lines bend toward it. the grid operator knows your name.
          "  "     " "     39 kWh burned for life  ·  2.3 kWh to level up

  DIET   █████████████░░░ 79% lean   cache ate 144 kWh it would've burned
         monstrous, yes — but it wastes almost nothing. the hog approves.

TOTAL ESTIMATE  ·  best guess, with the band it really sits in
  Energy  ≈ 39 kWh     14 kWh ▏░░░░░░░░░░░░░●░░░░░░░░░░░░▏ 109 kWh
          ≈ 16 hot showers  ·  391 pots of coffee  ·  112 hours on a gaming PC
  CO₂e    ≈ 15.6 kg    5.5 kg ▏░░░░░░░░░░░░░●░░░░░░░░░░░░▏ 43.8 kg  @ 400 g/kWh
          ≈ 6.8 liters of petrol burned  ·  130 km in a petrol car  ·  4.7 beef burgers
          a mature tree would need 284 days to breathe it back in
  Water   ≈ 43.0 L    data center cooling

Numbers mean nothing until you can feel them. 39 kWh is an abstraction. 130 km in a petrol car is a hog that's been eating well. Watthog turns every figure into something physical, and ships it with the band it really sits in — because the honest answer to "how much did my AI burn?" is a range, not a digit.

🔥 What a run actually costs

The same compute, in units you can picture. (Illustrative, grounded in the anchors in Methodology — your real mix decides the rest.)

| If watthog says… | …that's roughly | |---|---| | 0.1 kWh — a heavy afternoon of agentic coding | a few slices of toast 🍞 | | 1 kWh — a big refactor across a repo | an hour on a gaming PC, or a load of laundry 🧺 | | 10 kWh — a month of daily Claude Code | 4 hot showers, or 100 km in a petrol car 🚗 | | 40 kWh — a power user's quarter | a tank's worth of espresso ☕ and ~280 tree-days of CO₂ to reabsorb 🌳 |

Watthog's equivalence engine picks the comparison that fits your number automatically — toast for the light days, beef burgers and petrol for the months you'd rather not think about.

🧠 The hog has a level — and a villain arc

Here's the thing watthog refuses to do: congratulate you for burning power. Your footprint isn't a high score. So instead of cheering, watthog evolves a mascot that gets gloriously, self-awarely monstrous as your cumulative lifetime energy climbs. Cute → cute-evil → eldritch. You did this.

   ⚡ LEVEL UP  ·  Lv 27 → Lv 28  ·  SUBSTATION SWINE ⚡
   the hog has evolved. there is no going back.

• 🪜 Infinite log scale. Each level costs 25% more lifetime energy than the last — it never caps. The heaviest users on public token leaderboards land around Lv 65, with headroom to spare.

• 🎭 15 named forms, each with its own ASCII art and color:

  Piglet → Hoglet → Hog → Chonk → Big Chonkus → The Unit → Substation Swine → Hognarök → Hogzilla → Megahog → Brimstone Sow → Infernal Boar → The Coal Baron → The Grid Eater → Singularity Swine

  In an interactive terminal the hog moves — type hog to watch it.

• 💾 It remembers. Your level lives in ~/.config/watthog/config.json, so crossing a threshold greets you with a ⚡ LEVEL UP ⚡ on the next run.

And the DIET line is the one number you can actually move: how much energy prompt caching spared you versus sending every token fresh. The level only ever climbs — you can't un-burn what you've spent — but a lean hog wastes nothing.

🔌 What it reads

Watthog scans logs already on your machine. Nothing is uploaded; the only network calls are to your own Cursor and Copilot accounts, when you opt in.

| Agent | Source | |---|---| | Claude Code | ~/.claude/projects/**/*.jsonl | | OpenCode | ~/.local/share/opencode/opencode.db (+ legacy storage/message/) | | Codex CLI | ~/.codex/sessions/**/rollout-*.jsonl | | Cursor | Your usage export from Cursor's dashboard API¹ (local fallback: state.vscdb) | | GitHub Copilot | Local chat sessions + your premium-request report from GitHub's billing API² |

Every run opens with a SOURCES panel — instant proof of what was found and what still needs a nudge:

SOURCES
  ✓ Claude Code   14,902 messages · 312 sessions
  ✓ Codex CLI     1,285 messages
  ⚠ Copilot       412 requests (local only)
                    → run `watthog connect copilot` to include billing usage
  · Cursor        not installed
  run `watthog doctor` to inspect paths and troubleshoot missing sources

¹ Cursor stopped storing token counts locally in early 2026, so watthog fetches your usage export from cursor.com — authenticated with the session token Cursor itself keeps on your machine. It talks only to your own account and caches for an hour. Auto-mode requests hide the model and are estimated as mid-size. Offline or signed out, it falls back to the local database (covers usage up to ~Jan 2026).

² Copilot keeps no token counts on your machine, so watthog counts requests and estimates tokens per request — rougher than the token-accurate sources. Requests come from VS Code's local chat sessions (the only trace of the free "included" models) and GitHub's billing API for premium usage (per model, last 12 months, covers other machines). Where they overlap, the higher count wins. Connect the billing half once:

watthog connect copilot

This runs GitHub's device flow — a short code, one Authorize click, done. No token to create or copy. The access token (scoped to nothing but reading your plan) is saved to ~/.config/watthog/config.json. Prefer your own token? A fine-grained PAT with Plan: read, WATTHOG_GITHUB_TOKEN, or a scoped gh login are all picked up automatically. Without any of it, local sessions still count.

🚀 How it works

npx watthog                  →  scan local logs, print the energy report
watthog connect copilot      →  connect Copilot's premium-request billing
watthog doctor               →  show which sources are detected + what they need

Options

--days <n>   Only include the last n days
--co2 <g>    Grid intensity in gCO2e/kWh (default 400 global avg; Sweden ≈ 30)
--json       Machine-readable output

Three steps: run it → read it → fix the wasteful part. (We'd rather you shrink the number than brag about it.)

📊 Methodology

Watthog maps each model to a size class with a Wh-per-1k-token factor, and always reports a low–high range around the median:

| Class | Examples | Wh / 1k output tokens | |---|---|---| | small | Haiku, *-mini, *-flash, Gemma | 0.03 | | medium | Sonnet, GPT-4o, Gemini Pro, ~70B open models | 0.19 | | large (frontier) | Opus, Fable, GPT-5, o3 | 0.45 |

Input tokens are weighted at 1/8 of output (prefill batches far better than decode), cache writes at the input rate, cache reads at 1/80. Reasoning tokens count as output. The uncertainty band is ×0.35–×2.8.

• Llama-3.3-70B on H100/FP8 measures ~0.39 J per output token (~0.11 Wh/1k GPU-only). Scaled to full data-center energy via Google's Gemini disclosure (accelerator = 58% of total; CPU/RAM 25%, idle failover 10%, PUE 8%) → ~0.19 Wh/1k for the medium class.
• AI Energy Score (Hugging Face, H100 benchmarks of 166 models) and How Hungry is AI? for the spread between size classes.
• EcoLogits for the API-consumer view of closed models.
• Google (2025): median Gemini text prompt = 0.24 Wh, 0.26 ml water — the source of the 1.1 ml/Wh water factor.
• CO₂e uses a configurable grid factor (default 400 gCO₂e/kWh ≈ global avg; --co2 30 for Sweden). Equivalents ground emissions in Mike Berners-Lee, How Bad Are Bananas? and EU fleet averages: ~120 gCO₂e/km petrol car, ~80 g/banana, ~55 g/day reabsorbed by a mature tree, ~2310 g/L petrol, ~3300 g/beef burger.

These are estimates. Real batching, quantization and hardware can shift per-token energy by an order of magnitude either way — which is exactly why every number ships with a range.

🤝 Contributing

This hog is open source and hungry for contributions. PRs, issues and ideas welcome.

• Commit messages follow Conventional Commits:

  feat: add support for Cursor cache-hit detection
  fix: handle missing .codex directory gracefully

• Branch protection is on main — open a PR, get a review, merge.
• By contributing you agree to license your work under MIT.

License

MIT © Boomboxbacardi — see LICENSE.