TokenScope

Know exactly how many tokens (and dollars) your AI coding tools are burning on your project.

TokenScope automatically tracks token usage and estimated cost from Claude Code and OpenAI Codex CLI per workspace in VSCode. No API keys, no manual input — just install and code.

Why TokenScope?

AI coding tools don't show cumulative token usage per project. If you're working across multiple repos, it's hard to know where your tokens are going — or how much it's costing you.

TokenScope answers that by reading local conversation logs and presenting a clear breakdown — right inside VSCode.

• How many tokens did I spend on this project today?
• Which model is eating the most tokens?
• How much is this costing me?
• Is my usage trending up or down?
• Am I spending more on Claude Code or Codex CLI?

Supported Tools

Features

Status Bar — Always Visible

A persistent token counter in the status bar shows your project's total usage and estimated cost at a glance. Click it to open the full dashboard.

The tooltip expands to show a full breakdown:

• Input / Output / Cache Write / Cache Read / Reasoning tokens
• Total session and response counts
• Estimated cost in USD

Sidebar Tree View

Browse usage broken down by model and session directly in the activity bar.

• By Model — see each model's total tokens, call count, and estimated cost
• Sessions — expand any session to see per-token-type breakdown, cost, and which provider (Claude Code or Codex CLI) it came from
• Configurable session display limit with +N more sessions indicator

Dashboard

A dedicated webview panel with rich visualizations:

• Provider indicator — shows which tools are being tracked
• Summary cards — total tokens, input, output, cache write, cache read, reasoning, session count, response count, and estimated cost
• Per-model bar chart — see which models are consuming the most tokens, with cost per model
• Daily trend chart — visualize your token usage over the last 14 days to spot patterns and spikes
• Session history table — recent sessions with provider tag, full token breakdown, and per-session cost estimate
• CSV Export — one-click export of all session data for reporting or analysis

Cost Estimation

TokenScope automatically estimates your spending based on known model pricing:

Claude Models:

OpenAI Models:

Cost estimates appear in the status bar, tree view, and dashboard. If a model is not in the pricing table, those tokens are excluded from the cost total and a note is shown.

You can toggle cost display on/off via the tokenScope.showCostEstimate setting.

CSV Export

Export your complete session history to CSV for spreadsheets, reporting, or team sharing.

Each row includes: Session ID, Provider, Date, Response count, Input/Output/Cache Write/Cache Read tokens, Total tokens, and Estimated Cost.

Two ways to export:

• Click the Export CSV button in the dashboard
• Run TokenScope: Export Usage to CSV from the Command Palette

Daily Trend Chart

The dashboard includes a 14-day bar chart showing daily token consumption. Use it to:

• Spot which days had heavy AI coding tool usage
• Track whether your token usage is increasing or decreasing
• Identify spikes and correlate them with development activity

Zero Configuration

TokenScope reads from local log directories that Claude Code and Codex CLI already write to. There's nothing to set up.

Live Updates

File watchers monitor both Claude Code and Codex CLI log files. New usage appears automatically as you work — no reload needed.

Diagnostic Logging

TokenScope writes to a dedicated Output Channel (TokenScope in the Output panel). If something looks wrong — missing data, unexpected counts — check the logs for details on file parsing and directory resolution.

How It Works

Claude Code

~/.claude/projects/-Users-you-project-name/
  ├── 1a2b3c.jsonl    ← Claude Code writes conversation logs here
  ├── 4d5e6f.jsonl
  └── ...

1. Claude Code stores conversation logs as JSONL files under ~/.claude/projects/, keyed by your workspace path.
2. TokenScope maps your current workspace folder to the matching Claude Code directory.
3. It parses every assistant response to extract input_tokens, output_tokens, cache_creation_input_tokens, and cache_read_input_tokens.

Codex CLI

~/.codex/sessions/
  ├── rollout-2025-05-07T17-24-21-<uuid>.jsonl
  ├── rollout-2025-05-08T09-12-33-<uuid>.jsonl
  └── ...

1. Codex CLI stores session rollouts as JSONL files under ~/.codex/sessions/.
2. Each file starts with a session_meta entry containing the workspace directory (cwd).
3. TokenScope scans all rollout files and matches them to your current workspace.
4. It extracts token_count events with input_tokens, output_tokens, cached_input_tokens, and reasoning_output_tokens.

Aggregation

Data from all matched sources is merged and aggregated by session, model, and day, then displayed in the status bar, tree view, and dashboard. Cost is estimated by matching the model name against known pricing.

Token Types

Commands

Tip: You can also click the token counter in the status bar to open the dashboard.

Settings

Requirements

• VSCode 1.85.0 or later
• Claude Code and/or Codex CLI must have been used in the workspace at least once (so that log files exist)

FAQ

Q: Does this send any data to external servers? A: No. TokenScope is fully offline. It only reads local JSONL files from your machine. No telemetry, no network requests.

Q: Does it work with Claude.ai (web) or ChatGPT? A: No. TokenScope only tracks usage from CLI/IDE tools that store logs locally — currently Claude Code and Codex CLI.

Q: I opened a project but it shows 0 tokens. A: Make sure you've used Claude Code or Codex CLI in that specific workspace folder at least once. Check the TokenScope Output Channel for diagnostic details.

Q: Are the cost estimates accurate? A: They are estimates based on publicly listed API pricing. Actual costs may differ depending on your plan or billing terms. If a model isn't recognized, its tokens won't be included in the cost calculation.

Q: Can I see more than 50 sessions? A: Yes. Change the tokenScope.sessionLimit setting to any value between 5 and 500.

Q: Does it work on Windows? A: Yes. TokenScope supports both Unix and Windows-style line endings in JSONL files.

Q: Will more tools be supported in the future? A: We'd like to! However, Gemini CLI, Cursor, and GitHub Copilot don't store token usage data locally, making them impossible to track without API access. If those tools add local logging in the future, we can add support.

Development

git clone https://github.com/hooni0918/TokenScope.git
cd TokenScope
npm install
npm run compile
# Press F5 in VSCode to launch Extension Development Host

License

MIT