tokenlens

See exactly what's eating your Claude Code context window — and stop it.

npx -y github:bouncei/tokenlens status

That's the install. No clone, no build, no config.

Why this exists

Claude Code's built-in /context is documented to undercount MCP overhead by ~40k tokens per session. Anthropic was asked to fix it in March 2026 and closed the issue as "not planned." That's the wedge: the vendor's interest is to keep token consumption opaque to maximize per-user revenue; yours is to see exactly where your $200/mo Claude Max budget goes.

tokenlens is aligned with your wallet, not theirs.

What you actually see

Run against my live session as I'm writing this:

$ tokenlens status
tokenlens status
  session  ~/.claude/projects/-Users-josh-tokenlens/abc.jsonl
  turns    278 assistant

Tokens consumed by Anthropic
  input (uncached)           38.9k
  cache writes               2.07M   ← bloat lives here
  cache reads               41.21M   ← cheap
  output                    298.4k
  cache hit rate             95.2%

Output by attributionSkill
  <unattributed>          ████████████████   208.7k   (213 turns)
  idea-scout              ██████░░░░░░░░░░    92.6k   (65 turns)

Injected context by source
  cache_invalidation_or_growth  ████████████████ ! 520.3k   across 16 turns
  deferred_tools_added          █░░░░░░░░░░░░░░░ ~  25.0k   3 ev, 0 B text
  skill_listing                 ░░░░░░░░░░░░░░░░     1.5k   1 ev, 13.3 KB text
  hook_additional_context       ░░░░░░░░░░░░░░░░      791   1 ev, 5.5 KB text

MCP servers (loaded vs. invoked)
  (~50.5k estimated on tools never called)
  7dfd9fcf…               ████████████ ~  10.3k   ● never called
  plugin_playwright_playw…███████░░░░░ ~   5.8k   ● never called
  playwright              ███████░░░░░ ~   5.8k   ● never called
  Claude_in_Chrome        ██████░░░░░░ ~   5.5k   ● never called
  filesystem              ████░░░░░░░░ ~   3.5k   ● never called
  memory                  ███░░░░░░░░░ ~   2.3k   ● never called
  ...
  202 dead tools (use --show-dead for full list)

Two findings worth the install on their own:

1. ~50,500 tokens of context overhead this session was spent on MCP tools that were never invoked. Every server I had loaded — filesystem, memory, playwright, all 10 of them — got definition-injected and never called. That's /context reporting "you're at 60%" while 25% of the load is dead weight you could disable with one line in .mcp.json.

2. 520k tokens were written to cache as part of TTL-expiry invalidations across 16 separate moments — roughly $4 of cache-write cost on Opus, paid silently across the session and invisible to /context. tokenlens names it as cache_invalidation_or_growth so you can see when it's happening.

Commands

tokenlens status

One-shot breakdown of the active session. Reports Anthropic's exact per-turn usage from message.usage, attribution by attributionSkill, and a share-by-weight breakdown of which injected sources contributed to cache writes. A separate cache_invalidation_or_growth bucket captures cache writes that can't reasonably be tied to any visible attachment.

Flags:

• --cwd <path> — resolve the active session for a different working directory
• --session <file> — read a specific .jsonl directly

tokenlens watch

Same view as status, refreshed live as the session log grows. Useful to keep open in a side terminal while you work.

tokenlens watch

Press ctrl-c to stop.

tokenlens doctor

Detects (and optionally fixes) the two pathologies most commonly responsible for hidden context growth per Issue #29971:

1. Stale plugin versions — ~/.claude/plugins/cache/<source>/<plugin>/ directories still holding previous versions of installed plugins.
2. Duplicate skill symlinks — ~/.claude/skills/ symlinks pointing to the same canonical target, doubling injection of the same skill.

Default is dry-run; pass --fix to actually remove. Every deletion is logged with full path, and the tool refuses to touch anything outside ~/.claude/.

tokenlens doctor          # report only
tokenlens doctor --fix    # actually delete

tokenlens init and tokenlens budget (v2 preview)

Configure a weekly token budget and check your current session against it:

tokenlens init --tier max5      # writes ~/.claude/tokenlens.json with sane defaults
tokenlens budget                # shows: ALLOW / WARN / ASK / DENY vs your budget

These are the building blocks for v2's PreToolUse hook — see "v2 preview" below.

v2 preview — proactive budget enforcement (testable today)

The next surface is a Claude Code plugin that runs on every tool call and:

• Silently allows if you're under 50% of your weekly cap
• Injects a warning into Claude's context at 50–80% (Claude can decide to wrap up early)
• Asks for confirmation at 80–95% (you click through)
• Denies at 95%+ (or earlier if hardCap: true in your config)

It's already wired up. Test it locally:

git clone https://github.com/bouncei/tokenlens.git
cd tokenlens
pnpm install                                # builds dist/
tokenlens init --tier max5                  # writes ~/.claude/tokenlens.json
claude --plugin-dir ./tokenlens-cc          # loads the plugin

Now every tool call hits the budget evaluator. Watch the decisions in ~/.claude/tokenlens.log. The PreToolUse hook contract is fail-allow — a misbehaving hook will never block your workflow.

Marketplace submission to claude-plugins-community is pending.

How it works

tokenlens reads ~/.claude/projects/<encoded-cwd>/<session-uuid>.jsonl. The session log already contains Anthropic's ground-truth per-turn token counts, so for past turns tokenlens reports exact numbers — not estimates.

Each line is a JSON object representing one event in the session. The relevant fields:

• type: "assistant" rows contain message.usage: Anthropic's exact per-turn token count (input, cache_creation, cache_read, output, including the ephemeral_5m vs ephemeral_1h cache split).
• attributionSkill on each assistant row identifies which skill Anthropic attributed the turn to.
• type: "attachment" rows include the literal content of injected skills, hook context, todo reminders, and MCP-server tool deltas.

We aggregate these, share-out each turn's cache_creation across the attachments that preceded it (weighted by tokenized content size, or a heuristic for text-less injections), and cap per-event attribution at 1.5× weight so cache invalidations don't get falsely pinned on whatever attachment happened to be most recent.

Full schema notes in docs/SCHEMA.md.

Install (other paths)

The fastest install is npx, but if you want a stable binary on $PATH:

# Globally
pnpm add -g github:bouncei/tokenlens
# or
npm install -g github:bouncei/tokenlens

# From source
git clone https://github.com/bouncei/tokenlens.git
cd tokenlens
pnpm install      # also builds dist/ via the prepare script
./dist/index.js status

Once published to npm:

npx tokenlens status

Requires Node.js 20 or newer.

Roadmap

• v1 (shipped, v0.1.0): CLI status, watch, doctor, init, budget. Free, OSS.
• v2 (in progress, testable now): tokenlens-cc plugin wiring tokenlens into Claude Code's PreToolUse hook — proactive budget warnings, hard caps, and session-context injection. Pending community marketplace review. Free tier stays free; paid tier ($9–19/mo) adds cross-machine sync, weekly digest, team views (not yet built).
• v3 (next 12 months): Same value prop, cross-IDE. Cursor, Cline, Gemini CLI.

See DESIGN.md for the architecture overview.

Caveats

• Reverse-engineered against Claude Code v2.1.138 (May 2026). The session jsonl schema may shift on subsequent versions; tokenlens preserves unknown fields and won't crash on them.
• The local tokenizer (@anthropic-ai/tokenizer) tracks the Claude 2 era. Estimates for content we tokenize locally are ±5% vs. server-side counts. Past-turn numbers from message.usage are exact.
• The cache_invalidation_or_growth bucket lumps together genuine cache TTL expirations and conversation growth — tokenlens can't yet distinguish them without more session signal. Working on it.

Contributing

Issues and PRs welcome. The project is small and the surface is well-bounded — start with docs/SCHEMA.md to understand the session jsonl format, then CONTRIBUTING.md for how to add a new attachment category or analyzer.

If your ~/.claude/ has a structure tokenlens doesn't recognize, please open an issue with a sample line (anonymized) — every new attachment type we map deepens the wedge.

License

MIT — see LICENSE.