claude-token-lens

English | 中文

Keep your Claude Code context window alive longer — monitor usage, compress inputs, and make every session go further.

claude-token-lens installs as a pair of Claude Code hooks that silently compress your prompts before they reach the API, and filter verbose tool output before Claude processes it. It also provides a CLI for session statistics and configuration.

Works on top of every official Claude Code token-saving feature — not instead of them.

---

Why

The real cost of wasted tokens isn't money — it's context window lifespan. Every bloated prompt pushes earlier conversation history out of the window sooner, forcing you to /compact or /clear more often and losing hard-won context.

Coding sessions are especially wasteful. The three biggest culprits:

Source	Example waste
Stack traces	200-line traceback when you only need the error + 5 frames
Pasted code	Comment-heavy files, excessive blank lines
Test/log output	10,000-line log file when you only care about ERROR lines

Claude Code already provides /compact, /clear, auto-compaction, and prompt caching. claude-token-lens handles the part that happens before the prompt is submitted — so those official tools have less to work with in the first place. Less input pressure → fewer forced compactions → longer sessions before you lose context.

---

Features

• UserPromptSubmit hook — compresses every prompt before submission:
  • Stack trace condensation (Python, Node.js, Java, Go, Rust)
  • Code block cleanup (blank lines, optionally strip comments)
  • Whitespace normalization
• PreToolUse hook — filters Bash tool output:
  • Test runners (pytest, npm test, go test, …) → failures only
  • Log/file reads → last N lines
• Session statistics — token-lens stats shows cumulative savings (last 30 days); token-lens stats session shows the current session
• Monthly cost tracking — token-lens stats month shows the total API-equivalent cost for the current calendar month across all sessions, compared against the Claude Pro plan ($20/mo), so you can see whether Pro is paying off
• Real-time status line — embeds a live counter in Claude Code's status bar: lens 8.4k↑ 2.1k↓ · ctx 15% · usage 20% · Resets in 3hr 23min · $0.0535 · saved 1.2k (14%)
• Context window indicator — status line turns yellow at 50% and red at 80% with a /compact or /clear prompt
• Zero config to start — token-lens setup writes the hooks into ~/.claude/settings.json
• Configurable — tune thresholds, enable comment stripping, adjust frame counts

---

Quick start

# 1. Install
pip install claude-token-lens

# 2. Register hooks with Claude Code
token-lens setup

# 3. Restart Claude Code — hooks are now active

That's it. Every prompt you submit is now automatically compressed when savings exceed the threshold (default: 20 tokens).

---

Installation from source

git clone https://github.com/Yaminie-Hsu/claude-token-lens.git
cd claude-token-lens
pip install -e ".[dev]"
token-lens setup

---

How it works

UserPromptSubmit hook

Claude Code calls the hook with the prompt JSON before sending to the API. The hook runs three compression passes in order:

prompt → [stack-trace] → [code-blocks] → [whitespace] → compressed prompt

Stack trace compression keeps the first N and last N frames, replacing the middle with a summary line:

Traceback (most recent call last):
  File "app.py", line 5, in <module>       ← kept (head)
    main()
  File "app.py", line 12, in main          ← kept (head)
    result = process(data)
  File "handlers/base.py", line 34, in process  ← kept (head)
    return self._dispatch(req)
    ... [7 frames omitted] ...              ← ← ← replaced
  File "utils/schema.py", line 55, in _get  ← kept (tail)
    return [f for f in self.fields ...]
  File "utils/schema.py", line 70, in _get  ← kept (tail)
    raise AttributeError("no fields defined")
AttributeError: no fields defined

Code block compression collapses consecutive blank lines inside fenced code blocks and optionally strips single-line comments.

Whitespace normalization collapses runs of 3+ blank lines in plain text to 2.

After compression, the hook prints a one-line report to stderr:

[token-lens] 562 → 517 tokens (-45, -8%)  [stack-trace, code-blocks, whitespace]

If compression is below the report_threshold_tokens (default 20), the original prompt is passed through silently.

PreToolUse hook

Intercepts Bash commands before Claude runs them:

# Original command Claude wanted to run:
pytest tests/ -v

# Hook rewrites it to:
(pytest tests/ -v) 2>&1 | grep -A 10 -E '(FAILED|ERROR|FAIL|...)' | head -150; exit ${PIPESTATUS[0]}

Only failures surface in Claude's context. The exit code is preserved so Claude still knows whether tests passed.

Status line

token-lens setup also registers a statusLine script in ~/.claude/settings.json. After every Claude reply, the status bar updates with live session totals:

lens  8.4k↑ 2.1k↓ · ctx 15% · usage 20% · Resets in 3hr 23min · $0.0535 · saved 1.2k (14%)

• ↑ input tokens, ↓ output tokens (cumulative for the session)
• ctx % — context window fill level (200k); turns yellow at 50%, turns red at 80% with a /compact or /clear prompt
• usage % — 5-hour rate limit consumption, same metric as the official Claude Code "Current Session" bar
• Resets in — time until the rate limit window resets
• $ — session cost in USD (persisted to a local DB on every refresh for monthly tracking)
• saved — only shown when compression has fired

When nothing has been saved yet, the saved part is omitted:

lens  8.4k↑ 2.1k↓ · ctx 15% · usage 20% · Resets in 3hr 23min · $0.0535

The counter resets with each new Claude Code session. The cost is automatically accumulated across sessions in a local SQLite database — no extra steps needed.

Monthly cost report

token-lens stats month

── token-lens  monthly cost (April 2026) ─────────────────
  Sessions recorded  : 23
  API-equivalent cost: $24.3100
  Claude Pro plan    : $20.00/month
  ─────────────────────────────────────────────────────
  Net value          : +$4.31  (Pro 划算，等效省了这么多)

The "API-equivalent cost" is what you would have paid on a pay-as-you-go API plan. Comparing it to the $20 Pro monthly fee tells you whether your subscription is pulling its weight. You can also pass a specific month:

token-lens stats month 2026-03

---

CLI reference

# Show savings report for the last 30 days (or N days)
token-lens stats
token-lens stats 7

# Show stats for the current session only
token-lens stats session

# Show monthly API-equivalent cost vs Claude Pro plan
token-lens stats month
token-lens stats month 2026-03

# Compress text from stdin, print result to stdout
echo "my prompt" | token-lens compress
echo "my prompt" | token-lens compress --strip-comments

# Install hooks into ~/.claude/settings.json
token-lens setup

# Show / create the config file
token-lens config

---

Configuration

After running token-lens setup, a config file is created at ~/.claude-token-lens/config.json:

{
  "compressor": {
    "stack_head_lines": 3,
    "stack_tail_lines": 5,
    "max_blank_lines_in_code": 1,
    "strip_code_comments": false,
    "large_code_warn_tokens": 300,
    "collapse_whitespace": true,
    "report_threshold_tokens": 20
  }
}

Key	Default	Description
stack_head_lines	3	Stack frames to keep from the top
stack_tail_lines	5	Stack frames to keep from the bottom
max_blank_lines_in_code	1	Max consecutive blank lines inside code blocks
strip_code_comments	false	Remove // and # comment lines from code
large_code_warn_tokens	300	Warn when an inline code block exceeds this size
collapse_whitespace	true	Collapse 3+ blank lines in plain text to 2
report_threshold_tokens	20	Min tokens saved before reporting

To enable comment stripping (biggest savings for comment-heavy code):

{ "compressor": { "strip_code_comments": true } }

---

Relation to official Claude Code features

claude-token-lens is designed to complement, not replace, the token-saving tools built into Claude Code. Here is how they fit together:

Layer	Tool	What it does
Before prompt	claude-token-lens UserPromptSubmit	Compress the prompt you're about to send
Before tool	claude-token-lens PreToolUse	Filter verbose Bash output
During session	/compact [instructions]	Summarize conversation history
During session	/clear	Start fresh between unrelated tasks
During session	/effort / MAX_THINKING_TOKENS	Reduce extended thinking budget
Model choice	/model sonnet / haiku for subagents	Use cheaper models for simpler work
Project config	CLAUDE.md under 200 lines + Skills	Keep base context small
MCP	/mcp to disable unused servers	Reduce tool definition overhead
Monitoring	/usage + status line	Track context window usage

---

Compatibility

The hook-based features (UserPromptSubmit and PreToolUse) require a local shell environment to execute. The document CLI commands (preprocess, outline, timeline) work independently of Claude Code in any terminal.

Environment	Hooks	Document CLI
Claude Code CLI (claude in terminal)	✅ Full support	✅
Claude Code Desktop App (Mac / Windows)	✅ Same engine, same settings.json	✅
VS Code / JetBrains extension	✅ Backed by Claude Code engine	✅
claude.ai/code (web)	❌ No local shell access	✅ Run separately in terminal

If you use the Desktop App or an IDE extension, send a prompt with a long stack trace after setup and check your terminal for a [token-lens] line to confirm hooks are active.

---

Requirements

• Python 3.9+
• Claude Code CLI (any version with hooks support)
• tiktoken (installed automatically; falls back to character-based estimation if unavailable)

---

Running tests

pip install -e ".[dev]"
pytest tests/ -v

---

Contributing

Issues and pull requests are welcome. Areas where contributions would be most useful:

• Additional stack trace formats (Ruby, PHP, Swift, Kotlin)
• Smarter comment detection (language-aware instead of pattern-based)
• Integration tests against real Claude Code sessions
• Windows path support in setup command

---

License

MIT