Claude Carbon Tracker

A VSCode extension to track and visualize carbon emissions from Claude Code AI usage.

Features

• Real-time tracking of carbon emissions from Claude Code usage
• Status bar widget showing current CO₂ emissions
• Detailed statistics view in the sidebar with:
  • Total CO₂ emissions
  • Token counts (input/output)
  • Request counts
  • Environmental impact equivalents
• Configurable emission factors to adjust calculations
• Persistent tracking across VSCode sessions

Installation

From Source

1. Clone this repository
2. Run npm install
3. Run npm run compile
4. Press F5 to open a new VSCode window with the extension loaded

From VSIX (once published)

1. Download the .vsix file
2. Run code --install-extension claude-carbon-tracker-0.0.1.vsix

Usage

Once installed, the extension automatically starts tracking Claude Code usage:

1. View in Status Bar: Look for the leaf icon (🌿) in the bottom-right status bar
2. Open Sidebar: Click the Claude Carbon Tracker icon in the Activity Bar
3. View Statistics: Click the status bar item or run Claude Carbon: Show Statistics
4. Reset Stats: Run Claude Carbon: Reset Statistics from the command palette

Configuration

Configure the extension in your VSCode settings:

{
  "claudeCarbonTracker.emissionFactor": 0.0004,
  "claudeCarbonTracker.showInStatusBar": true
}

• emissionFactor: CO₂ emissions in kg per 1,000 tokens (default: 0.0004)
• showInStatusBar: Show/hide the status bar widget

How It Works

Data Collection Methodology

The extension monitors Claude Code's local conversation data stored in JSONL files:

1. File System Monitoring: Scans Claude Code's data directory every 5 seconds

   • Windows: %USERPROFILE%\.claude\projects\
   • macOS/Linux: ~/.claude/projects/ or ~/.config/claude/projects/

2. Token Extraction: Parses conversation files to extract:

   • input_tokens: Tokens sent to the model
   • output_tokens: Tokens generated by the model
   • Cache tokens are tracked but not counted (following Claude's billing model)

3. Carbon Calculation: Applies emission factors to token counts

   • Formula: CO₂ (kg) = (total_tokens / 1000) × emission_factor
   • Supports configurable emission factors for different scenarios

4. Privacy-First: All processing happens locally; no data leaves your machine

Emission Factor Methodology

The default emission factor (0.0004 kg CO₂/1K tokens) is derived from recent research on LLM carbon footprint:

Calculation Basis

Energy Consumption:

• Research estimates 3-4 joules per token for large language models [1]
• For Claude 3.7 Sonnet: ~0.4 joule/token [2]
• Conversion: 1 kWh = 3,600,000 joules
• Energy per 1K tokens: (400 joules × 1000) / 3,600,000 ≈ 0.11 Wh ≈ 0.00011 kWh

Carbon Intensity:

• GPT-4 operational emissions: ~0.3 gCO₂e per 1K tokens (Azure US West, 240.6 gCO₂e/kWh) [3]
• Claude reports 3.5 gCO₂ per query [4]
• Global data center average: ~400 gCO₂eq/kWh [5]
• Conservative estimate: 0.4 gCO₂/1K tokens = 0.0004 kg CO₂/1K tokens

Factors Considered:

• Server hardware power consumption (GPU/CPU TDP)
• Power Usage Effectiveness (PUE) of data centers (typically 1.2-1.5×)
• Grid carbon intensity (varies by region: 57-429 gCO₂eq/kWh) [5]
• Inference efficiency (Claude 3.7 Sonnet is noted as highly eco-efficient) [6]

Model-Specific Considerations

Different Claude models have different computational requirements:

• Haiku: Faster, more efficient, lower emissions per token
• Sonnet: Balanced performance and efficiency (current default)
• Opus: More computational power, higher emissions per token

Note: This extension currently uses a single emission factor. Future versions will support model-specific factors.

Environmental Impact Equivalents

The extension converts CO₂ emissions to relatable comparisons:

• Trees needed for 1 year: CO₂ (kg) / 21 kg [7]
• Kilometers driven: CO₂ (kg) / 0.12 kg/km (average car) [8]
• Smartphones charged: CO₂ (kg) / 0.011 kg [9]
• 60W light bulb hours: CO₂ (kg) / 0.0006 kg/hour [10]

Limitations and Assumptions

1. Inference Only: Does not account for model training emissions (one-time cost amortized across users)
2. Network Overhead: Does not include data transmission emissions
3. Regional Variation: Uses average carbon intensity; actual emissions vary by data center location
4. Cache Benefits: Claude's prompt caching reduces actual emissions but is not fully modeled
5. Hardware Diversity: Assumes consistent hardware; actual GPU/CPU efficiency varies

Uncertainty Range

Based on literature review, estimated emissions could range from:

• Lower bound: 0.0002 kg CO₂/1K tokens (optimal conditions: renewable energy, efficient hardware)
• Default: 0.0004 kg CO₂/1K tokens (conservative average)
• Upper bound: 0.001 kg CO₂/1K tokens (worst case: coal-powered grid, inefficient hardware)

Users can adjust the emission factor in settings based on their preferred assumptions or updated research.

Roadmap

• [x] Monitor Claude Code's local conversation files
• [x] Real-time token tracking and carbon calculation
• [ ] Support for different models (Sonnet, Opus, Haiku) with model-specific emission factors
• [ ] Detect model type from conversation files
• [ ] Export statistics to CSV/JSON
• [ ] Daily/weekly/monthly reports and trends
• [ ] Carbon offset recommendations
• [ ] Integration with carbon offset APIs
• [ ] Historical emissions visualization charts

References

[1] Luccioni, A. S., et al. (2023). "Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model." Journal of Machine Learning Research, 24(253), 1-15. https://jmlr.org/papers/volume24/23-0069/23-0069.pdf

[2] FOSS Force. (2025). "What's Your Chatbot's Carbon Footprint?" Retrieved from https://fossforce.com/2025/04/whats-your-chatbots-carbon-footprint/

[3] arXiv:2507.11417v1. (2025). "Quantifying the Energy Consumption and Carbon Emissions of LLM Inference via Simulations." https://arxiv.org/html/2507.11417v1

[4] Fast Company. (2025). "The environmental impact of LLMs: Here's how OpenAI, DeepSeek, and Anthropic stack up." https://www.fastcompany.com/91336991/openai-anthropic-deepseek-ai-models-environmental-impact

[5] Li, Y. L., et al. (2024). "Towards Carbon-efficient LLM Life Cycle." HotCarbon '24: Workshop on Sustainable Computer Systems. https://hotcarbon.org/assets/2024/pdf/hotcarbon24-final154.pdf

[6] Clune, J. (2025). "Environmental Impact of AI." Arthur's Blog. https://clune.org/posts/environmental-impact-of-ai/

[7] U.S. Environmental Protection Agency. "Greenhouse Gas Equivalencies Calculator." https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator

[8] European Environment Agency. (2024). "CO2 emissions from passenger transport." https://www.eea.europa.eu/

[9] U.S. Department of Energy. "Energy Use Calculator." https://www.energy.gov/

[10] Carbon Footprint Ltd. "Carbon Footprint Calculator." https://www.carbonfootprint.com/

Additional Reading

• Frontiers in Communication (2025). "Energy costs of communicating with AI." https://www.frontiersin.org/journals/communication/articles/10.3389/fcomm.2025.1572947/full

• Springer Nature. (2024). "Green AI: exploring carbon footprints, mitigation strategies, and trade offs in large language model training." Discover Artificial Intelligence. https://link.springer.com/article/10.1007/s44163-024-00149-w

• GitHub greenscale-ai/genai-carbon-footprint. "GenAI Carbon Footprint Calculator." https://github.com/greenscale-ai/genai-carbon-footprint

• Stanford CRFM. (2024). "Anthropic: Claude 3 - Foundation Model Transparency Index." https://crfm.stanford.edu/fmti/May-2024/company-reports/Anthropic_Claude%203.html

Contributing

Contributions welcome! This is an early-stage project focusing on:

• Refining emission factor calculations with latest research
• Supporting model-specific emission factors
• Improving accuracy of environmental impact estimates
• Enhancing UI/UX and data visualization
• Adding export and reporting features

Please cite relevant research when proposing changes to emission factors.

License

ISC

Disclaimer

This extension provides estimates only based on available research and reasonable assumptions. Actual carbon emissions may vary significantly based on:

• Data center location and energy sources: Anthropic's infrastructure location and renewable energy usage
• Model architecture and optimization: Hardware efficiency, quantization, and other optimizations
• Network infrastructure: Data transmission emissions not included in current calculations
• Caching and efficiency measures: Claude's prompt caching and other efficiency features
• Temporal variations: Time of day, grid mix changes, and seasonal factors
• Hardware specifics: GPU/CPU models, utilization rates, and cooling requirements

Research Limitations

Current research on LLM carbon emissions is evolving, and many companies (including Anthropic) provide limited transparency about their environmental impact. The emission factors used in this extension are derived from:

• Peer-reviewed academic research
• Industry reports and benchmarks
• Reasonable engineering estimates

For authoritative data, refer to:

• Anthropic's official sustainability reports (when available)
• Peer-reviewed research on LLM carbon footprint
• Third-party audits and certifications

Educational Purpose

This tool is designed to raise awareness about the environmental impact of AI usage and encourage more sustainable practices. It should not be used for:

• Carbon accounting for regulatory compliance
• Official carbon offset calculations
• Comparative claims without proper context

We encourage users to critically evaluate the assumptions and stay informed about latest research in this rapidly evolving field.