Token Aware

A VS Code / Cursor extension that keeps your AI token/dollar usage visible in the status bar in real time, so you stay aware of how much you're burning without opening a usage dashboard.

It auto-detects which tools you use (Cursor, Claude Code, or both) and shows one status bar item per detected tool. Tools that aren't on your machine simply don't appear.

What it does

• One status bar item per detected tool:
  • ⚡ $1.93 session · $0.42 last · $1301.81 mo · 62% (Cursor)
  • ✨ 45.2k session · 922 last · 43% session · 9% week (Claude)
• Auto plan detection: dollars if you're billed per usage, tokens if you're on a flat monthly plan. Signing in with corporate SSO under an organization's per-usage contract is detected too, so you never need to hold an API key yourself. Manual override if it guesses wrong.
• Per-session spend: tracks the session you're working in. It becomes "current" the moment you send a message in it.
• The cost of your message, not your last API call: one reply is often many round trips to the model. "Last" covers everything since you hit enter, so it answers "what did that cost me" instead of flickering between sub-calls.
• Where the cost actually went: hover for the split between answering your message, loading context in, and re-reading context already loaded. On a long conversation the last one is usually the biggest share, which is the part most people don't expect.
• Plan-limit gauges (Claude subscription plans): the session and weekly percentages from claude.ai → Settings → Usage, with warning colors at 90% and 100%.
• Updates on a poll interval and instantly after each AI turn.
• Click for a details panel: current session, spend by session, monthly totals, top models, and a per-model breakdown of the current session.
• Claude session list is scoped to your open workspace by default. Monthly totals still count everything.

How it works

Cursor

Cursor doesn't store real per-request token counts locally (the chat DB records zeros), so the numbers come from Cursor's own usage backend, the same data behind cursor.com/dashboard/usage:

1. Reads your local session token from Cursor's state DB (ItemTable key cursorAuth/accessToken in state.vscdb), read-only, via the system sqlite3 binary. Nothing is written to that DB.
2. Calls Cursor's usage endpoints with that token for per-request token usage, cost, and plan type.
3. Watches ~/.cursor/projects/*/agent-transcripts/<id>/*.jsonl to know which agent is active and to refresh immediately after each turn.

Claude Code

Claude Code writes full transcripts to ~/.claude/projects/<slug>/<sessionId>.jsonl, and every assistant message includes an exact usage block. So this provider is fully local, accurate, and needs no auth: it reads those files, dedupes streamed chunks by requestId, groups by session, resolves session titles, and (optionally) converts tokens to dollars with a bundled Anthropic pricing table. Plan detection reads the local Claude Code login metadata (billingType and plan tier); a claude.ai login by itself isn't treated as proof of flat-fee billing, since Team/Enterprise seats sign in the same way but are commonly billed per-usage. Nothing leaves your machine for this check.

Note: neither editor exposes the focused chat tab to extensions, so "current session" means the conversation you most recently interacted with (it updates as soon as you send in it).

Requirements

• Cursor tracking: be signed in to Cursor; sqlite3 on PATH (preinstalled on macOS; on Windows: winget install SQLite.SQLite).
• Claude tracking: Claude Code installed (data under ~/.claude/projects). No auth needed.

Develop / run

npm install
npm run build      # or: npm run watch

Open this folder in VS Code/Cursor and press F5 to launch the Extension Development Host.

Package / install locally

npm run package    # produces token-aware-<version>.vsix

Install the .vsix via the Extensions view → ... → Install from VSIX...

Settings

Commands

• Token Aware: Refresh Now
• Token Aware: Show Details

Privacy

Everything runs locally. The Cursor provider only talks to cursor.com (the same host Cursor itself uses). The Claude provider reads transcripts entirely locally; for subscription accounts it makes one authenticated call to api.anthropic.com (the same endpoint Claude Code's own /usage screen uses) to show plan-limit percentages. No third-party telemetry. Auth tokens are read at runtime and never stored or transmitted anywhere except to the tool's own backend.

Notes / disclaimers

• Not affiliated with or endorsed by Anysphere (Cursor) or Anthropic (Claude).
• Cursor's usage endpoints (src/cursorApi.ts) are unofficial and may change; the client fails soft (shows "n/a" instead of erroring).
• Claude dollar amounts come from a bundled table of Anthropic list prices (src/claudePricing.ts), matched per model version and dated 2026-07-22; token counts are exact. Fast mode is priced from usage.speed, so /fast calls bill at their real premium rather than being assumed away. Two modifiers are not modelled because nothing in a transcript records them, and neither applies to Claude Code: the Batch API's 50% discount and the inference_geo: "us" 1.1x. Run npm test to re-derive every published rate from the table.
• Plan auto-detection is heuristic (e.g. an enterprise seat without usage-based billing is still shown in dollars); use the unit settings to override.

License

MIT