claude-token-lens: see exactly what's burning your Claude quota, live

What it is

claude-token-lens is an npm tool that runs a live terminal dashboard alongside Claude Code. It reads the JSONL session files Claude Code writes to ~/.claude/projects/ and shows you — turn by turn, in real time — which tool, agent, skill, or MCP server is burning your token quota.

When it actually helps

If you have ever hit a quota wall mid-project and had no idea whether to blame a chatty MCP server, a recursion-happy skill, or your own sprawling prompts, this fills the gap. The attribution table breaks your 5-hour window into per-source rows: each tool call, each subagent spawn, each skill invocation gets its own line with an output-token count and a percentage of total.

The live command runs in a second terminal while Claude Code runs in the first. No login, no telemetry, nothing leaves your machine. Once you can see which source is eating the most tokens, you have something concrete to act on: tighten a prompt, swap a tool, or drop a skill.

The honest caveat

Anthropic does not publish exact quota limits. Every number claude-token-lens displays is built on community-reverse-engineered estimates — Pro ~44k output tokens per 5-hour window, Max5 ~88k, Max20 ~220k. These are directionally useful, not authoritative. Claude Code's built-in /stats command can show different numbers because of org-level pooling, server-side cache invalidation, and an undisclosed quota formula.

The weekly caps Anthropic added in August 2025 are not tracked yet either. If you are on a plan with weekly limits, this tool will not warn you about the slower-burning ceiling. Use it to understand relative burn — "MCP server X is taking 40% of my tokens" — not to predict exactly when you will be throttled.

npx claude-token-lens live

Built by wassimbensalem — source and full docs at github.com/wassimbensalem/claude-token-lens. MIT license.