UNDERSTAND → TRY → COMPARE → KEEP OR UNDO
Reading usage from this computer…
These are log stores, not accounts. Shared launchers count once. Custom folders outside your home need an explicit path. Missing data is not zero usage. Unresolved-event counts cover all retained logs; response and legacy counts use the selected period.
Check the source status above, expand the time range, or point the app at your custom log directories. Use Claude Code or Codex and keep this dashboard open.
tracefrugal watch --claude-dir /path/to/claude --codex-dir /path/to/codex
Cloud sessions, Claude Desktop, and ChatGPT conversations are not imported.
Check your installed Claude version against official docs, back up one project's settings, and try a smaller MCP result limit. A fresh-session receipt confirms activation. Compare hourly usage and answer quality, then keep or restore.
tracefrugal optimize --project /path/to/your/projectLocal optimizer setup and verified versions →
It joins instructions, available tool definitions and retained history.
The code and arguments it generates are output for this request.
The result becomes input to the next model request. Retained earlier answers also become input.
| What you see | How it is counted |
|---|---|
| A generated answer, code or tool-call arguments | Model output when generated; input again if retained in a later request. |
| A browser Result, file contents or command result | Tool-returned data supplied as model input. It is not model-generated output. |
| A terminal diff, progress log or truncated preview | Screen formatting does not establish billed tokens. Use provider usage receipts. |
Illustration only: 100k input per request × request count. Real context grows, is trimmed or compacted. Cache hits lower the applicable rate but still count as input; reuse is not guaranteed.
Reduce oversized tool results and unnecessary round trips first. Preserve useful cached context. These logs do not measure the exact share of your question, instructions, schemas and history.
Official tool-calling flow ↗ · Official prompt-cache accounting ↗Recorded input, cache, and output tokens.
Choose one small instruction to test. The app does not change your model or reasoning effort.
Choose a session to see how each request grew.
Responses include tool loops and subagents. Tokens are the total processed across requests, including repeated context. Select a row for its identity, input distribution and cost formula.
03 · COMPARE TOKENS AND SATISFACTION
The previous 24 hours vs. up to 24 hours after a change. Hourly observations and your ratings stay on this computer.
GET STARTED
For people who use Claude Code or Codex on a Mac, Windows PC, or Linux computer.
Download and extract the archive for your operating system. No Go, Python, API key, or sign-up is needed.
Download TraceFrugal ↗macOS: arm64 for Apple Silicon, amd64 for Intel.
Windows: amd64 for most PCs, arm64 for ARM PCs.
Linux: choose the matching processor architecture.
In a terminal opened in the extracted folder:
./tracefrugal
The app opens your browser, finds local Claude Code and Codex stores, and refreshes automatically. Keep its terminal running.
If a browser does not open, use the printed local URL. On macOS, if the downloaded app is blocked, review it in Privacy & Security and choose Open Anyway only if you trust the download. The release is not notarized.
| Source | How it works |
|---|---|
| Claude Code | Local JSONL discovery in ~/.claude and ~/.claude-*; also CLAUDE_CONFIG_DIR. Main sessions and recorded subagents. |
| Codex | Local sessions and archived_sessions in ~/.codex and ~/.codex-*; also CODEX_HOME. Request receipts preferred; older snapshots have explicit coverage limits. Native Codex costs remain unpriced. |
| Claude Code with Bedrock or another provider | Tokens when recorded in the same format. Provider-specific model IDs may remain unpriced. |
| OpenAI / Anthropic API applications | Available through the API recorder and evaluator. Requires integration. |
| ChatGPT, Claude Desktop, web conversations | No native import. |
Prices are token-based estimates, not a subscription bill or a plan-limit meter. Your provider's billing page remains authoritative.
Coverage, accounting, privacy and custom sources →