green-code

Track your AI carbon footprint and plant trees to compensate.

What it does

• Shows a carbon footprint summary at every Claude Code session start (CO2 emitted, trees planted, debt, progress bar towards next tree)
• Silently tracks your Claude Code token consumption at each session end
• Estimates energy use (kWh) and CO2 emissions based on published research
• Two modes: auto (plant trees automatically at threshold) or manual (track and plant on demand)
• Plants real trees via Tree-Nation API

Install

/plugin marketplace add jsamson/green-code
/plugin install green-code@green-code-marketplace

On first launch, the plugin asks for your Tree-Nation API token and preferences.

Commands

Command	Description
/green:status	View your carbon footprint and tree balance
/green:plant	Plant trees to offset accumulated CO2
/green:plant 3	Plant a specific number of trees
/green:config	View current configuration
/green:config mode auto	Switch to auto-plant mode

How it works

Token tracking

At each session end, the plugin reads ~/.claude/stats-cache.json and computes the delta since the last check.

Energy estimation

Per-model base cost in Wh per output token. Other token types are derived as ratios that match Anthropic's API pricing tiers (which track underlying compute cost):

Model family	Wh / output token	Rationale
Opus	0.002	Dense ~400B-class, TokenPowerBench Llama-405B reference
Sonnet	0.0008	Mid-size MoE / dense
Haiku	0.0003	Small fast model
(unknown)	0.001	Conservative default

Token type	Ratio vs output	Effective Wh (Opus)
Output	1.00	0.002
Input (fresh)	0.20	0.0004
Cache creation	0.25	0.0005
Cache read	0.02	0.00004

CO2 conversion

• Default: 320 gCO2/kWh (cloud-provider weighted US mix with renewable PPAs; AWS sustainability report 2024)
• PUE 1.15 for datacenter overhead, aligned with AWS 2024 global average
• 1 tree = 10 kg CO2 (configurable)

Tree-Nation

Trees are planted via the Tree-Nation REST API. You need a Tree-Nation account and API token.

Sources

• TokenPowerBench, AAAI 2026 -- arxiv.org/abs/2512.03024 (per-model output-token energy at frontier scale)
• "Energy Considerations of LLM Inference", arXiv 2504.17674 (prefill vs decode energy)
• Luccioni et al., "Power Hungry Processing", ACM FAccT 2024
• AWS Sustainability Report 2024 -- aws.amazon.com/sustainability/data-centers/ (PUE 1.15)
• Anthropic Prompt Caching docs (cache read ~90% reduction vs fresh input)
• IEA, "Electricity 2024" -- iea.org/reports/electricity-2024

Localization

Shell script output (session start, setup wizard, planting feedback) is auto-translated based on $LANG. Supported locales: fr and en (default). Skills (/green:status, /green:plant, /green:config) remain in English -- they are Markdown prompts loaded by Claude, who will naturally respond in the user's language anyway.

Limitations

• Energy estimates have ~x2 uncertainty -- no AI provider publishes per-token energy data
• Claude's exact architecture (dense vs MoE) is not public
• CO2 depends on datacenter location; US mix is the default assumption
• stats-cache.json may not update in real-time in all Claude Code versions

Prerequisites

• jq (JSON processor) -- install: sudo apt install jq / brew install jq
• bc (calculator) -- usually pre-installed on Linux/macOS
• curl -- usually pre-installed
• Tree-Nation account + API token

License

MIT