Agent Ledger
Every token your AI coding agents burn, in one ledger.
Install it and it finds the AI coding agents you use. It shows every token they've burned (GitHub-style calendar, cost, sessions, models, projects), read from the logs those tools already keep on your computer. No API keys, no proxy, nothing uploaded.
Formerly API Health Monitor. v1 tracked direct API calls through a local proxy; Agent Ledger tracks coding agents from their own logs instead. v1 code is kept in v1/.

| Sidebar |
Click any day |
 |
 |
🔍 Supported agents
The extension auto-detects all of these on first run and tells you what it found.
| Agent |
Support |
What's read |
| Claude Code (CLI, VS Code, JetBrains, desktop Code tab) |
✅ Full |
~/.claude/projects/**/*.jsonl |
| Codex (CLI, IDE extension, desktop app) |
✅ Full + rate limits |
~/.codex/sessions/**/rollout-*.jsonl[.zst] |
| Cline |
✅ Full + tool-reported cost |
<editor>/User/globalStorage/saoudrizwan.claude-dev/tasks |
| Roo Code |
✅ Full + tool-reported cost |
…/globalStorage/rooveterinaryinc.roo-cline/tasks |
| Kilo Code |
✅ Full + tool-reported cost |
…/globalStorage/kilocode.kilo-code/tasks |
| Gemini CLI |
✅ Full |
~/.gemini/tmp/*/chats/*.jsonl |
| Antigravity |
✅ Full (newer conversations) |
~/.gemini/antigravity*/conversations/*.db |
| OpenCode |
✅ Full + tool-reported cost |
~/.local/share/opencode (SQLite or JSON) |
| GitHub Copilot Chat |
⚠️ Partial |
<editor>/User/workspaceStorage/*/chatSessions |
| Cursor, Windsurf, Kiro, Trae, Zed, Amp, Copilot CLI, Factory Droid, Qwen Code, Continue, Augment, Goose |
ℹ️ Detected |
Shown as installed, with the reason their usage can't be counted |
Extensions such as Cline and Copilot are found in every VS Code-based editor: VS Code, Insiders, VSCodium, Cursor, Windsurf, Antigravity, Trae, Kiro and Positron, plus Remote/WSL server installs. Works on Windows, macOS and Linux.
Why some tools are "detected only": Cursor, Windsurf, Amp and Trae keep usage on their servers, not on your computer. Inventing numbers for them would be worse than saying so. Agents you run inside those editors (Claude Code, Codex, Cline, Copilot) are still counted.
Why Copilot is "partial": Copilot Chat records the prompt size of only the last step of each request, plus total output. In agent mode, input is therefore a lower bound. Copilot bills per premium request rather than per token, so no dollar cost is estimated for it.
✨ Features
|
|
| 🟩 Usage calendar |
GitHub-style contribution graph of tokens or cost per day, for 6 or 12 months or any year. Click a day for its sessions, tools and hourly breakdown. Shows current streak, longest streak and busiest day. |
| 📊 Overview |
Tokens, cost, sessions, active days, cache hit rate, average per request and a month-end projection |
| 🧰 By tool |
Every agent side by side. Click a chip to filter. |
| 📈 Activity |
Daily or weekly stacked chart, a weekday × hour heatmap and totals by weekday |
| 💬 Sessions |
Every session with its title, project, requests, active time, models and cost. Searchable and sortable. |
| 🤖 Models / 📁 Projects |
Where your tokens and money went |
| ⏱ Limits |
Codex rate-limit windows, plus "API value you got for your plan" for each tool |
| 🔌 Agents |
What was found where, what can't be read and why. Ignore an agent with one click. |
| 📌 Status bar |
Today's tokens and cost |
| 📤 CSV export |
Every request as a row |


💰 How cost is calculated
- Reported cost. When the tool logs its own cost (Cline, Roo Code, Kilo Code, OpenCode), that number is used.
- Estimated cost. Otherwise, tokens are priced at public API list prices (Nov 2026), including cache reads and writes. If you're on a subscription you actually pay your plan price, so set
aiUsage.subscriptions to see how much value you get from it.
- Not billed per token. Copilot shows "billed per request".
Override any model's price with aiUsage.pricing.
✅ Accuracy
Each parser has tests built from real log samples (npm test). On the author's machine, totals were checked against independent recounts:
- Claude Code: exact match on every token field.
- Codex: exact match with Codex's own per-session totals.
- Cline: exact match with Cline's own per-task totals.
Other details:
- Responses that appear in more than one file are counted once.
- Claude Code deletes transcripts after 30 days. This extension keeps the usage it has already read, so your history keeps growing.
⚙️ Settings
| Setting |
Default |
|
aiUsage.refreshIntervalMinutes |
5 |
Rescan interval. Unchanged files are skipped, so rescans are cheap. |
aiUsage.statusBar |
true |
Today's usage in the status bar |
aiUsage.subscriptions |
{} |
Monthly plan price per agent id, e.g. { "claude-code": 100, "codex": 8 } |
aiUsage.pricing |
{} |
Price overrides (USD per 1M tokens), e.g. { "gpt-5.6-luna": { "input": 0.2, "output": 1.2, "cacheRead": 0.02 } } |
aiUsage.disabledAgents |
[] |
Agent ids to ignore |
aiUsage.extraPaths |
{} |
Extra data folders per agent id, for unusual installs |
Agent ids: claude-code, codex, copilot-chat, cline, roo-code, kilo-code, gemini-cli, antigravity, opencode.
Commands: Rescan Logs, Open Full Dashboard, Export Usage as CSV, Show Scan Log, Clear Cache and Rescan Everything.
🔒 Privacy
The extension only reads the files listed above and makes no network requests. It stores a cache of token counts (not your prompts) in VS Code's extension storage. Session titles come from the tools' own titles, or from the first line of your first prompt when a tool has none.
🧩 Adding an agent
Each agent is a self-contained provider:
export interface Provider {
id: string; name: string; color: string;
support: 'full' | 'partial' | 'detect-only';
version: number; // bump to re-parse cached files
discover(ctx): Promise<{ roots; files; notes? }>; // find data folders + files
parse?(file): Promise<FileResult>; // file -> usage records
enrich?(ctx, roots, sessions, records): Promise<void>; // titles/projects from index files
}
- Add
src/providers/<agent>.ts.
- Register it in
src/providers/index.ts.
- Add a test with a fixture in
src/test/.
The scanner handles caching, parallel parsing, de-duplication and history retention.
🚀 Development
npm install
npm test # compile + 20 unit tests
node out/smoke.js # scan this machine's real logs from the terminal
npx @vscode/vsce package
src/
extension.ts activation, webviews, status bar, commands
core/ scanner (cache, concurrency), payload, pricing, paths, sqlite, protobuf
providers/ one file per agent + index.ts registry
test/ node:test unit tests with generated fixtures
media/ dashboard UI (vanilla JS + SVG, no libraries)
License
MIT