Skip to content

inspiration.md

Latest commit

 

History

History
58 lines (39 loc) · 2.14 KB

inspiration.md

File metadata and controls

58 lines (39 loc) · 2.14 KB
Raw

Inspiration: phuryn/claude-usage

Source: https://github.com/phuryn/claude-usage

What it does

A local dashboard that reads Claude Code's JSONL session transcripts and shows token usage, cost estimates, and session history. Works on API, Pro, and Max plans.

Tech stack

  • Python 3.8+, standard library only (no pip install, no venv)
  • sqlite3 — persistent storage at ~/.claude/usage.db
  • http.server — serves a single-page HTML/JS dashboard
  • Chart.js on the frontend for visualizations

Core files

  • scanner.py — parses JSONL into SQLite, incremental (tracks file mtime)
  • dashboard.py — HTTP server, single-page UI
  • cli.py — command dispatcher

CLI commands

  • python cli.py scan — populate the DB from JSONL files
  • python cli.py today — today's breakdown by model (terminal)
  • python cli.py stats — all-time stats (terminal)
  • python cli.py dashboard — scan + open browser at localhost:8080
  • Env vars: HOST, PORT, --projects-dir for custom paths

What it captures

  • Claude Code CLI (claude in terminal)
  • VS Code extension (Claude Code sidebar)
  • Dispatched Code sessions

What it does NOT capture

  • Cowork sessions — server-side, no local JSONL
  • Cost calculations for non-standard model names (only matches opus/sonnet/haiku — others show n/a)

Data shape

Each JSONL line is a message. Usage lives at message.usage:

  • input tokens, output tokens
  • cache creation tokens, cache read tokens

Model identifier at message.model.

Limitations worth addressing in our version

  1. Cost for Pro/Max users is misleading — API rates shown, but those users pay a flat subscription. The original admits this in its README.
  2. Unknown models show n/a — no fallback pricing, no way to add custom pricing.
  3. UI is a single HTML page with Chart.js — functional but dated.
  4. No session drill-down — can't easily click into a session and see what happened.
  5. No per-project comparison — aggregates are global.
  6. 30-second auto-refresh only; no live tailing of the active session.

These are candidates (not commitments) for what our version could do better.