Claude Token Counter

Live token usage, cost, and context-window tracking for Claude Code, right in VS Code — a status bar item plus a sidebar dashboard with per-project history.

How it works

Claude Code has no public extension API for reading its session state. The only officially documented, stable data source for live usage/cost is its statusLine mechanism: Claude Code runs a script you configure and pipes it a JSON payload (tokens, cost, context %, rate limits, model) on stdin after every turn.

This extension:

1. Generates a small standalone Node.js script into its own storage folder.
2. Configures that script as Claude Code's statusLine command (workspace- or user-level settings.json).
3. The script appends every update it receives to an append-only log.
4. The extension tails that log, updates the VS Code status bar and sidebar in near real time, and rolls totals up into a local history file grouped by project.

The catch: statusLine is a terminal-only feature

statusLine only fires when Claude Code is rendering an interactive terminal UI. Sessions started via the official Claude Code VS Code extension's own chat panel run claude headlessly (--no-chrome --output-format stream-json) — there's no terminal to draw a status line into, so it never triggers, and this extension would otherwise show nothing for that entire (very common) way of using Claude Code in VS Code.

To cover that case, the extension also tails Claude Code's transcript files (~/.claude/projects/<project>/<session-id>.jsonl) as a best-effort fallback, extracting the same per-turn usage data Claude Code writes there regardless of UI mode. This format is explicitly undocumented by Anthropic and can change between releases, so:

• It's used only for sessions that never sent a statusLine update — a session already tracked via statusLine is left alone.
• Cost is estimated from token counts at list pricing (shown with ~ and "estimated" in the UI) rather than reported by Claude Code itself, since the transcript doesn't include a running cost total.
• It can be turned off via claudeTokenCounter.enableTranscriptFallback if a future Claude Code release breaks the parsing (everything is read defensively and fails silent-skip, never crashes, but the numbers could go stale or wrong).

Setup

On first activation the extension offers to enable itself. You can also run these commands manually (Cmd+Shift+P / Ctrl+Shift+P):

• Claude Token Counter: Enable for This Workspace — writes to .claude/settings.local.json in the open folder (not committed to git by Claude Code convention).
• Claude Token Counter: Enable Globally (All Projects) — writes to ~/.claude/settings.json, so every project gets tracked.

If you already have a custom statusLine configured, the extension asks before replacing it.

After enabling, start (or restart) Claude Code in a terminal in that workspace — the status bar updates after Claude's first response.

What it shows

• Status bar: model, session cost, and context-window usage (configurable via claudeTokenCounter.statusBarFormat).
• Sidebar panel (Claude Token Counter icon in the Activity Bar):
  • Current session — tokens, cache read/write, context % for whichever session is active in this workspace.
  • Plan limits (Anthropic) — your Pro/Max 5-hour and 7-day quota usage, with reset times. This is the account-level limit Anthropic enforces, as opposed to what you've actually spent. It's a last-known snapshot (kept visible even between sessions) rather than tied to the current session, since statusLine is the only source for it and doesn't fire for every session — see the limitation below.
  • Your usage — today's and the last-7-days' totals across every project, with a daily bar chart. Built from a delta-based day-by-day rollup, so it's accurate even across sessions that span midnight.
  • Projects (all time) — cumulative cost/tokens per project.

Cost is whatever Claude Code itself reports (cost.total_cost_usd) for statusLine-tracked sessions — computed client-side at list rates, so it may differ slightly from your actual invoice. Transcript-fallback sessions get an estimate instead (see above).

Limitations

• Plan limits (5h/7d quota %) require a statusLine session — they're Anthropic account data that only rides along on statusLine updates, so they're never available for chat-panel (transcript-fallback) sessions no matter what. Run Claude Code in a terminal at least occasionally to keep this section current.
• Rate-limit percentages (5h / 7d) only appear for Claude.ai Pro/Max subscribers to begin with; not for API-key/console billing.
• Requires node to be resolvable from the shell Claude Code runs in (true for any machine that has Claude Code installed).
• Windows without Git Bash: statusLine commands run through PowerShell instead of bash; the generated node "<path>" command should still work, but this hasn't been tested there.

Development

npm install
npm run watch   # esbuild in watch mode

Press F5 in VS Code to launch an Extension Development Host with the extension loaded.