🪨 TokenCave

One extension to rule your tokens.

Token counter · Usage bars · Cache timer · Caveman mode · Context compressor

---

What is TokenCave?

TokenCave is a browser extension for claude.ai that combines two things:

1. Token intelligence — see exactly how much context you're using, when it expires, and how close you are to limits
2. Caveman mode — slash your output tokens by ~75% by making Claude speak with radical brevity — without losing technical accuracy

---

Features

📊 Token Counter

Approximate token count for your current conversation, with a mini progress bar against Claude's 200k context limit. Uses the same o200k_base tokenizer family.

⏱ Cache Timer

Live countdown showing how long your conversation remains cached. Cached messages cost significantly less — this timer tells you when to keep going vs. when to start fresh.

📈 Usage Bars

Session (5-hour) and weekly (7-day) usage with progress bars and reset countdowns. More accurate than Claude's /usage page because it reads exact utilization fractions from the SSE stream, not rounded percentages.

🪨 Caveman Mode

Make Claude respond like a brilliant, extremely terse caveman. Same technical substance — roughly 75% fewer output tokens.

Level	Example
🪶 Lite	Drop filler and hedging, keep full sentences
🪨 Full	Drop articles, use fragments, short synonyms
🔥 Ultra	Abbreviate everything, arrows for causality (X→Y), one word when sufficient

Commands (type in chat):

/cave          → toggle on/off
/cave lite     → switch to lite mode
/cave full     → switch to full mode
/cave ultra    → switch to ultra mode
/cave off      → disable
/cave status   → show current state

Or use the extension popup to click between levels.

🗜 Context Compressor (CLI)

Compress your CLAUDE.md, project notes, and documentation files to cut input tokens by ~46% — without losing meaning. Runs locally via Python.

pip install -r requirements.txt
python -m skills.compress.scripts path/to/CLAUDE.md

---

Installation

Chrome / Edge / Chromium

1. Download the latest release ZIP (or clone this repo)
2. Go to chrome://extensions and enable Developer mode
3. Drag and drop the ZIP onto the page (or click Load unpacked)

Firefox

1. Go to about:debugging#/runtime/this-firefox
2. Click Load Temporary Add-on → select manifest.json

---

Before / After — Caveman Mode

🗣 Normal Claude (69 tokens) "The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object."

🪨 Caveman Claude · full (19 tokens) "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."

Same fix. ~75% fewer tokens. Brain still big.

---

Privacy & Security

• ✅ All data stays in your browser — no external servers, no tracking
• ✅ Only communicates with claude.ai
• ✅ Reads lastActiveOrg cookie only to query Claude's own /usage endpoint
• ✅ postMessage restricted to window.location.origin (not '*')
• ✅ Caveman state stored in sessionStorage — clears when you close the tab
• ✅ Compression scripts refuse to process secrets, .env, keys, or credentials

See SECURITY.md for full details.

---

Caveman Mode — How It Works

When caveman mode is active, TokenCave prepends a bracketed instruction to your message before it's sent:

[RESPONSE STYLE: full caveman — drop articles, fragments OK, short synonyms…]

Your actual question here

Claude respects this instruction and responds tersely for the rest of the message. The prefix is invisible in the UI — you just see the compressed response.

The instruction is NOT stored anywhere, NOT sent to any server, and clears when you turn caveman mode off.

---

Context Compressor — How It Works

The compression CLI (skills/compress/scripts/) uses Claude's API to:

1. Detect whether a file is natural language (compressible) vs. code/config (skip)
2. Refuse to process any file that looks like it contains secrets or credentials
3. Compress natural-language files to ~50% of original tokens
4. Validate the output preserves all technical content
5. Retry up to 2 times if quality check fails

Before: CLAUDE.md — 8,200 tokens
After:  CLAUDE.md — 4,400 tokens  (~46% reduction)

---

Project Structure

TokenCave/
├── manifest.json          # Extension manifest (MV3)
├── popup.html             # Extension popup UI
├── src/
│   ├── content/
│   │   ├── constants.js   # Shared constants & colors
│   │   ├── bridge-client.js
│   │   ├── tokens.js      # Token counting & caching
│   │   ├── caveman.js     # 🆕 Caveman mode manager
│   │   ├── ui.js          # UI components & DOM injection
│   │   └── main.js        # Orchestrator
│   ├── injected/
│   │   └── bridge.js      # Page-world fetch interceptor
│   ├── styles.css
│   └── vendor/
│       └── o200k_base.js  # Bundled tokenizer
├── icons/                 # Extension icons
├── skills/
│   ├── caveman/SKILL.md   # AI coding assistant skill (Claude Code / Cursor)
│   └── compress/
│       ├── SKILL.md
│       └── scripts/       # Python compression CLI
├── requirements.txt
├── SECURITY.md
├── CONTRIBUTING.md
└── LICENSE

---

Acknowledgements

• Token counting via gpt-tokenizer (MIT)
• Caveman speech compression concept — viral LLM observation that terse prompts preserve accuracy

---

License

MIT