Inspiration
AI queries feel free, but they are not. Each request consumes energy and water for cooling, even when the question could be answered instantly by a search engine. Many everyday prompts, e.g., checking a time zone, defining a concept, testing a string, which do not really need a large model. WattsUp was built to make this hidden cost visible and gently encourage more thoughtful use without restricting access to AI.
What it does
WattsUp is a Chrome extension that intercepts prompts on ChatGPT, Gemini, and Claude before they are sent. A local classifier categorizes the prompt and decides what to do with it. Simple factual lookups are redirected to Google Search, empty or low-value prompts trigger a small warning, and reasoning-heavy prompts start a short “think first” timer. At the same time, a live meter estimates tokens, energy use, and CO₂ for each request, adjusted for the selected model. A dashboard tracks total water and energy saved, weekly trends, and topic statistics, while a Supabase-powered leaderboard lets friends compare impact through shareable codes.
How we built it
The extension is written entirely in JavaScript with no frameworks or external libraries. It runs on Chrome Manifest V3 using content scripts for each supported AI site and a background service worker. Prompt classification uses a two-tier system: a fast regex layer with over forty patterns, followed by a small weighted scoring model for prompts. Also, we implement Supabase REST endpoints handle leaderboard synchronization.
Challenges we ran into
The challenge we had was classification accuracy, it required careful tuning, as overly strict rules blocked useful prompts, while loose rules reduced the system’s impact.
Accomplishments that we're proud of
- Supports multiple languages through pattern-based detection and includes a visual savings banner when prompts are redirected to search.
- The leaderboard works with live Supabase synchronization, allowing real comparisons between users.
- Rule-based systems can work surprisingly well when tuned with real usage data.
What we learned
Even small AI interactions have measurable environmental costs at scale. Many prompts people send to large language models could be handled more efficiently with traditional tools.
What's next for WattsUp
Next steps include supporting other search engines (e.g., Firefox, Bing), expanding the leaderboard into team-level dashboards, and improving classification with a lightweight ML model. We also want to explore carbon offset integrations so savings translate into real environmental contributions, and add developer-focused tracking to monitor token usage across APIs and projects.
Log in or sign up for Devpost to join the conversation.