burncard

Accurate local AI usage stats for Claude Code and Codex. Your token burn, your streak, your card.

npx burncard

CLAUDE CODE
  model     Fable 5  (Max 20x)
  tokens    9.72B
  messages  31.4K
  sessions  141
  streak    33d (longest 33d)
  active    49 days, peak 11 PM

Why another usage tool?

Because the obvious way to count is wrong:

• Claude Code streams each assistant message as multiple log entries that repeat the same usage block. Naive counters inflate your tokens ~2.4x. burncard dedupes by message and request id.
• Log timestamps are UTC. Tools that bucket days in UTC split your late-night sessions across two days and break your real streak. burncard uses your local timezone.
• Plan info lives in two local records that go stale independently. burncard trusts the freshest one, and only ever claims your tier — never a price it can't verify.

Privacy

Everything runs locally. burncard reads your own log files (~/.claude/projects, ~/.codex/sessions), computes aggregates, and prints them. Nothing leaves your machine.

Usage

npx burncard                    # stat block
npx burncard --json             # machine-readable
npx burncard card               # share card -> burncard.svg
npx burncard card --codex -o me.svg

As a library

burncard is a telemetry layer first — the CLI is just one consumer. Any app (status bars, overlays, dashboards, presence daemons) can read the same stats programmatically:

import { computeClaude, computeCodex, detectClaudePlan } from "burncard";

const stats = computeClaude(); // ToolStats | null
const plan = detectClaudePlan(); // "Max 20x" | "Pro" | ... | null

Roadmap

• Share card (SVG) — post your burn
• PNG export
• Cursor + Gemini CLI parsers
• burncard watch — Discord Rich Presence daemon

License

MIT