DEV Community

Cover image for Why I built EcoPrompt: A zero-telemetry Chrome extension to measure LLM compute costs
IB2M_Official
IB2M_Official

Posted on

Why I built EcoPrompt: A zero-telemetry Chrome extension to measure LLM compute costs

#webdev #ai #javascript #chromeextension

GenAI interfaces like ChatGPT, Google Gemini, and Claude are clean and instant. But behind every prompt lies power-dense GPU clusters drawing electricity and evaporative cooling water.

To bring physical transparency to AI interactions without tracking users or slowing down the browser, I built EcoPrompt — an open-source, 100% client-side Chrome extension (Manifest V3).

EcoPrompt Analytics Dashboard

🌿 What EcoPrompt Tracks

When you hit Enter on ChatGPT, Gemini, or Claude, EcoPrompt displays a non-intrusive in-page toast estimating:

  • 🍃 Carbon Footprint (g CO₂e): Real-time emissions based on datacenter PUE and energy grid averages.
  • 💧 Water Consumption (mL): Direct evaporative cooling plus indirect electricity water draw.
  • ⚡ Active Energy (Wh): Compute energy required for prompt prefill and decode cycles.
  • 💶 API Cost Equivalence: Baseline token computation rates across model tiers.

🛠️ Technical Constraints & Architecture Choices

Building a extension for developers required strict design decisions:

1. Zero Telemetry & 100% Client-Side Privacy

  • No prompt logging: Prompt text is never stored, read, or transmitted.
  • In-memory evaluation: Character counts are held temporarily in memory to apply KV-cache weightings, then immediately purged.
  • Local persistence: All analytics stay local via chrome.storage.local.

2. Pure Vanilla Architecture (< 150 KB)

  • Built with zero external dependencies (no npm bloat).
  • Dynamic timelines (Day, Week, Month, Year) run on a custom-built, lightweight native SVG charting engine taking under 10 KB.

3. Daily Quotas & Smart Reminders

  • Set soft daily targets to encourage prompt batching and context pruning without intrusive popups.

🚀 Try It Out

EcoPrompt is completely free and open-source under CC BY-NC-SA 4.0.

I'd love to get your feedback on the client-side modeling, SVG engine, or suggestions for new providers to support!

Top comments (1)

Collapse
 
ib2mofficial profile image
IB2M_Official •

Thanks for reading! 👋

I'm curious to hear your thoughts: do you currently factor in compute/API cost or energy consumption when choosing between models (e.g., Haiku/Flash vs. Opus/o1)?

Also, if you test out EcoPrompt, I’d love your feedback on the toast UI or performance! 🚀