AI Wattch - Track Your AI Footprint

Discover how your AI usage impacts the planet. Measure, compare, and optimize your AI footprint in real time.

Available on:

1. Chrome Web Store
2. Firefox Store

Project Summary

AI Wattch is an open-source browser extension powered byAntarctica’s One Token Model (OTM) that estimates the energy use and carbon footprint of end-user interactions with LLM-powered chat interfaces such as ChatGPT and Claude. It combines token-based and time-based estimation, regional infrastructure mapping, and model-specific parameters to deliver transparent, science-backed emissions reporting per session.

Why AI Wattch

• Transparency: Makes invisible energy costs visible - per session, per token, per model.
• Efficiency: Helps people prompt more efficiently and choose more efficient models.
• Privacy: Prioritizes privacy; no chat text leaves the browser.
• Scalability: Built for extensibility: multi-model, multi-region, multi-browser roadmap.

How it works (high level)

1. The extension parses the page DOM for supported chat UIs (ChatGPT, Claude).
2. It captures lightweight telemetry (timestamps, token counts, model selection) - never full chat text.
3. Two estimation approaches run (token-based & time-based). Both can be combined into a hybrid estimate.
4. Regional factors (PUE, grid carbon intensity) and model-specific hardware stats (TDP, quantization) convert energy → emissions.
5. The UI surfaces real-time metrics, session summaries, model comparisons, and prompt-efficiency tips.

Supported Models & Platforms

Initially supported (V1.5 / V2.0 planned mapping):

• ChatGPT model family (manual selection for ChatGPT Pro/Plus; automatic detection for free-tier where possible)
• Claude family (automatic detection via DOM parsing)
• Planned: Gemini, other LLMs (modular architecture supports adding new detectors)

Browsers

• Chrome (MV3) - current release
• Firefox - current release

Methodologies (overview)

AI Wattch runs two complementary estimators and a hybrid orchestration:

I. Token-based estimator (DOM-derived token proxy)

• Counts characters/tokens from DOM, converts to tokens (default 4 chars/token, configurable 3–5).
• Uses token energy factors (input/output) and infrastructure multipliers (PUE, grid intensity) to compute energy and emissions.

II. Time-based estimator (timestamp-derived compute duration)

• Uses timestamps T1 (request), T2 (first token), T3 (last token) to derive computation time.
• Maps computation time to GPU power, utilization, server baseline, and PUE → energy → emissions.

III. Hybrid & model-aware approach (Antarctica enhancements)

• Uses model-specific metadata (total/active params, quantization, estimated active GPUs, GPU TDP & memory, token generation rate) to refine active GPU count and per-token energy.
• Dynamically applies regional PUE and carbon intensity based on IP region or manual selection.
• Handles edge cases (summarization, streaming vs. batch, cached responses fallback logic).

---

Install / Build / Run (developer)

Prerequisites

• Node.js (LTS recommended)
• npm or yarn
• Chrome (for load-unpacked development)

Clone & Install

git clone https://github.com/AIWattch/AI-Wattch.git
cd ai-wattch-ext
npm install

Development Build (watch)

npm run dev

Load dist/ via chrome://extensions → Developer mode → Load unpacked → select browser-extension/dist.

Production Build

npm run build:extension

Artifact appears in dist/. Use packaged release workflow for Chrome Web Store.

Release

1. Create a release branch release/vX.Y.Z
2. Update CHANGELOG.md and package.json version
3. npm run build:extension → create zip → upload to GitHub Release and Chrome Web Store.

---

Development Workflow & Recommended Practices

• Language: TypeScript (strict mode)
• Linting: ESLint + Prettier (pre-commit hook)
• Commit style: Conventional commits (feat/fix/chore/docs)
• Branching: feature branches, one feature per PR
• PR checklist (required): build passes, tests pass, docs updated, minimal surface area

Testing

Unit tests: Vitest Run tests:

npm test

Tests cover:

• Token counting & character → token mapping
• Timestamp-based computation logic
• Basic model-detection flows (mock DOM)
• Regional lookup fallbacks

Integration & Scenario Tests (recommended):

End-to-end simulation with recorded DOM payloads (place in tests/fixtures/)

Contributing & Governance

We welcome contributions. Please follow these steps:

1. Fork the repo.
2. Create branch feature/<short-description>.
3. Commit with a clear message; open a PR.
4. Add tests and update docs.
5. One feature per PR; link relevant issue.

Pre-PR: For major architectural or methodology changes, open an issue to discuss design and data assumptions (methodology is research-sensitive). Maintain transparency in how variables are chosen and cite sources in PR descriptions.

Docs to Add/Maintain

• CONTRIBUTING.md (detailed)
• CODE_OF_CONDUCT.md
• SECURITY.md (vulnerability reporting)

Privacy & Security

• AI Wattch does not send chat contents off-device.
• The extension collects minimal telemetry (token counts, timestamps, model id) used only for computation.
• IP-based region detection is optional - users can manually set the region (privacy-first).
• For any detected security/privacy issue: follow SECURITY.md and do not open a public issue; contact maintainers.

Short Glossary & Variable Origin

• EcoLogits: baseline token energy literature.
• Artificial Analysis: latency & generation rate estimates.
• ArXiv: academic sources used for deriving token → latency relations and GPU utilization assumptions.
• Quantization Q: bytes per parameter mapping (INT4=0.5, INT8=1, FP16=2, FP32=4). Overhead factor ~1.2 applied.

Detection & Model Identification (implementation notes)

• Claude: DOM parsing available - use robust selectors and feature flags; test extensively against different Claude UIs.
• ChatGPT: free-tier detection possible via DOM; paid tiers restrict automatic detection - provide a manual model dropdown and clear UX to set model if detection fails.
• Fallbacks: assume global average PUE & carbon intensity if location/model cannot be resolved.

Example Test Cases (to include under tests/fixtures)

• Short QA exchange (50 tokens output): verify token and time estimators align within tolerance.
• Long summarization (15,000-word doc): ensure chunking recommendation triggers and energy spike is reported.
• Rapid retries (3 prompts within 3 minutes): ensure repetitive/iterative category detection and nudge.

License & Credits

• License: MIT (see LICENSE file).
• Built by Antarctica & IT Climate Ed with contributions from the open-source community. See AUTHORS.md.