AI Token Usage Tracker

A VS Code extension that analyzes GitHub Copilot (Chat and CLI) and Claude Code token usage, session statistics, and cost insights from local logs. Track how many tokens you consume, estimate costs against your plan quota, and drill into individual requests — all without sending data anywhere. Switch between providers with a single click.

Installation

Requires VS Code 1.85 or newer. Pick one of the two options.

Option 1 — Install the prebuilt .vsix (quickest)

1. Download the latest package from the releases/ folder of this repository — currently ai-token-usage-tracker-0.2.10.vsix.
2. Install it, either:
   • From the UI: open the Extensions view (Ctrl+Shift+X), click the ... menu → Install from VSIX..., and pick the file; or
   • From a terminal:

     code --install-extension ai-token-usage-tracker-0.2.10.vsix
     

3. Reload VS Code when prompted, then follow Setup.

To update later, install the newer .vsix the same way; it replaces the old version.

Installed an earlier build named "AI Usage Tracker" (local.copilot-usage-tracker)? Uninstall it first. The extension now has a new ID, so the two would run side by side. The new one starts with an empty cache and rebuilds its statistics from the logs still on disk.

Option 2 — Build from source

Requires Node.js and Git.

git clone https://github.com/alexandrulupei/Ai-Usage-Tracker.git
cd Ai-Usage-Tracker
npm install
npm run package        # compiles and writes releases/ai-token-usage-tracker-<version>.vsix
code --install-extension releases/ai-token-usage-tracker-0.2.10.vsix

To debug instead of installing, open the folder in VS Code and press F5 to launch an Extension Development Host. For a rebuild-on-save dev loop, run npm run watch.

Setup

GitHub Copilot

Copilot only writes the per-request logs this extension reads when agent debug file logging is on. Add this to your user settings.json (Ctrl+Shift+P → Preferences: Open User Settings (JSON)):

{
  "github.copilot.chat.agentDebugLog.fileLogging.enabled": true
}

Then reload the window (Ctrl+Shift+P → Developer: Reload Window) and run a Copilot chat.

Upgrading from an older setup? Earlier versions of this README also listed github.copilot.chat.agentDebugLog.enabled. That key no longer exists in current Copilot Chat (verified against Copilot Chat 0.73.0 in VS Code 1.145). VS Code ignores it, so you can delete it.

Notes:

• Chats from before you enabled logging are partly recovered from VS Code's chat history (see copilotUsageTracker.chatHistoryFallback). Credits are exact for those requests, but input tokens cover only the final round, so those sessions are marked incomplete.
• Other editors: logs are discovered automatically across VS Code-family editors (Code, Insiders, VSCodium, Cursor, Windsurf, …). JetBrains IDEs are detected on a best-effort basis; there, turn on agent debug file logging in the GitHub Copilot plugin's chat settings. Use Select IDE Sources to choose which editors to include.
• Copilot CLI: sessions from the standalone copilot CLI and VS Code's "Copilot CLI" chat type are read from ~/.copilot/session-state and need no extra setup.

Claude Code

No setup. Claude Code writes its session transcripts to ~/.claude/projects automatically.

Features

• Two providers, one view — Track GitHub Copilot and Claude Code side by side; toggle the active provider with the Switch Provider button (orange accent for Claude, blue for Copilot).
• Token Usage Dashboard — Input, output, cached (read), and cache-write tokens per session, model, or billing period. A "fresh in" figure leaves out the re-read cached prefix, so it can be compared with each tool's own usage panel.
• Cost estimation —
  • Copilot: AI credits / USD against your plan quota (Free, Pro, Pro+, Business, Enterprise). It uses the billed amount Copilot logs for each request, and list prices when no billed amount is logged. It also shows activity (sessions and chat requests).
  • Claude: USD at Anthropic list prices, with correct cache-read (0.1×) and cache-write pricing, including the 1-hour vs 5-minute TTL distinction.
• Session Explorer — Browse every conversation in a sidebar tree, grouped by date, workspace, or model. Each one has an expandable per-request timeline (prompt, output, tool inputs/results).
• Multi-IDE sources — Combines Copilot logs from every detected editor plus the Copilot CLI; include or exclude each with Select IDE Sources.
• Accurate request accounting — For Claude, the lines of one API response are de-duplicated to a single request by requestId. Deeply nested subagent/workflow transcripts are rolled into the parent session.
• Background refresh — Syncs at startup and then every 15 minutes by default (configurable).
• Billing-period navigation — Step through months or view all-time totals; the current billing period is always selectable.
• Offline & private — All processing is local. The extension only reads existing log files; it never intercepts traffic or phones home.

Usage

1. Open the AI Usage view in the Activity Bar (left sidebar).
2. The extension automatically discovers Copilot logs (all workspaces and detected editors) and Claude Code transcripts.
3. Use the Switch Provider button (view title bar) to toggle between Copilot and Claude.
4. Click Refresh to scan for new sessions immediately.
5. Open the Dashboard (AI Usage: Open AI Usage Dashboard) for charts, per-model breakdowns, and the expandable request timeline.
6. Use the billing-period picker to view a specific month or all-time totals.

Commands

Configuration

How it works

The extension reads existing local files written by each tool:

GitHub Copilot (when a session exists in the OTel DB, that copy wins)

1. OTel Agent Traces DB — SQLite spans with cached/reasoning tokens, exported via Copilot's exportAgentTracesDB command. Only populated when github.copilot.chat.otel.dbSpanExporter.enabled is on.
2. JSONL debug logs — workspaceStorage/*/GitHub.copilot-chat/debug-logs/, written when agent debug file logging is on.
3. Copilot CLI sessions — ~/.copilot/session-state/<id>/events.jsonl.
4. VS Code chat history — fallback for requests not covered above.
5. Prompt export — cached-token enrichment via Copilot's internal export (manual refresh only).

Claude Code

• Reads ~/.claude/projects/<project>/<sessionId>.jsonl and its nested subagent/workflow transcripts. One API response spans several transcript lines that share a requestId; these are de-duplicated to a single request so tokens are counted once. Cache-read, 5-minute cache-write, and 1-hour cache-write tokens are priced separately.

Costs are computed by a shared per-token engine keyed off the provider (not the model name), so a Claude model run through Copilot is still billed under Copilot's rules.

See DATA-COLLECTION.md and COST-CALCULATION.md for the full data-collection and billing methodology.

Data privacy

• No network calls — The extension never sends data to any server.
• Read-only sources — It only reads existing log/transcript files; it never modifies them.
• Local storage only — All statistics are cached in a local SQLite database in the extension's storage folder.
• Full content is stored locally — To power the expandable request timeline, the local database stores full message text, assistant output, and tool inputs/results (capped at 16 KB per field). This data stays on your machine and is removed by Clear Cache. If you don't want it stored, avoid installing this build on shared machines.

Development

npm run watch     # Compile in watch mode (dev)
npm run compile   # Production build into dist/
npm run package   # Build a .vsix into releases/

When you release a new version, bump version in package.json, run npm run package, commit the new .vsix in releases/, and update the download link under Installation.

License

MIT