🔢 Gemini Counter v2.0.0

A lightweight Chrome extension that shows exact token counts, a live context bar, output tokens, and model info — right inside gemini.google.com.

---

✨ Features

• Exact token counts — pulled directly from Gemini's own usageMetadata, not a guess
• Live context bar — visual progress bar with tiered warning states (75% → orange, 90%+ → red)
• Output tokens — shows N out alongside prompt token count
• Threshold markers — subtle tick marks at 75% and 90% on the context bar
• Dynamic tooltip — live % context usage on hover
• Conversation history — token counts persisted per-conversation via chrome.storage.local
• Model info — detects and displays the active Gemini model
• Fallback tokenizer — uses a local o200k_base estimate before the first real response arrives

---

📸 Screenshot

The counter widget appears in Gemini's input bar, showing estimated input tokens, a progress bar, and the active model — all without leaving the page.

---

🚀 Installation

Works on Chrome, Edge, and Arc (any Chromium-based browser).

1. Download this repo — click Code → Download ZIP or git clone
2. Open chrome://extensions in your browser
3. Enable Developer mode (toggle in the top-right)
4. Click Load unpacked and select the gemini_counter_v2 folder
5. Navigate to gemini.google.com — the counter appears automatically ✅

---

🛠 How It Works

Gemini's web app uses internal, obfuscated endpoints (unlike Claude's clean REST API). This extension works around that by:

1. Injecting a bridge script into the page context to intercept fetch and XMLHttpRequest
2. Parsing JSON response bodies to find usageMetadata — the same data Google's official API returns
3. Extracting promptTokenCount, candidatesTokenCount, and thoughtsTokenCount directly
4. Falling back to a local tokenizer estimate until the first real API response arrives

Token counts come straight from Gemini — so they're exact, not estimated.

---

🔒 Privacy

• All data stays local — no external servers, no analytics, no tracking
• Only reads Gemini's own response data from your browser session
• Makes zero additional network requests
• Token history stored only in chrome.storage.local

---

📐 Context Window Limits

Model	Context Window
Gemini 2.5 Pro	1,048,576 tokens (1M)
Gemini 2.5 Flash	1,048,576 tokens (1M)
Gemini 2.5 Flash Lite	1,048,576 tokens (1M)
Gemini 2.0 Flash	1,048,576 tokens (1M)
Gemini 1.5 Pro	2,097,152 tokens (2M)
Gemini 1.5 Flash	1,048,576 tokens (1M)

Unknown models fall back to 1M and log a warning to the browser console.

---

📁 Project Structure

gemini_counter_v2/
├── manifest.json               # Extension manifest (MV3)
├── icons/
│   ├── icon16.png
│   ├── icon48.png
│   └── icon128.png
├── assets/
│   └── screenshot.png          # UI screenshot
└── src/
    ├── styles.css              # Counter widget styles
    ├── vendor/
    │   └── o200k_base.js       # Local tokenizer fallback
    ├── injected/
    │   └── bridge.js           # Page-context fetch interceptor
    └── content/
        ├── constants.js        # Model limits & config
        ├── bridge-client.js    # Messaging between contexts
        ├── tokens.js           # Token counting logic
        ├── ui.js               # Widget rendering
        └── main.js             # Entry point & orchestration

---

🆕 What's New in v2.0.0

Bug Fixes

• UTF-8 corruption fix — TextDecoder now uses a single decoder with correct { stream: true } flag, preventing multi-byte character corruption at chunk boundaries
• postMessage origin hardened — All window.postMessage calls target window.location.origin instead of '*'
• Ghost duplicate widget fixed — attach() removes the container from its current parent before re-inserting
• DOM observer debounced — MutationObserver debounced 200ms to avoid layout thrashing on Gemini's SPA
• Dead hash API removed — Unused kind: 'hash' handler removed from bridge.js
• accumulatedText dead variable removed — Cleaned up unused stream handler variable

New Features

• Output token display (N out)
• Tiered warning states (normal → orange at 75% → red at 90%+)
• Bar threshold markers at 75% and 90%
• Dynamic tooltip with live % context usage
• Unknown model console warnings
• Conversation history storage (last 50 entries per conversation)
• gemini-2.5-flash-lite added to known models

---

🤝 Contributing

Contributions are welcome! Please read CONTRIBUTING.md before opening a PR.

---

📄 License

MIT — free to use, modify, and distribute.