tokens-gone

Local web dashboard for AI coding agent token usage and cost. Reads Claude Code, Codex, OpenCode V1 and V2, Pi, and Oh My Pi session logs, aggregates tokens and cost, and lets you slice the data interactively.

Run

npx tokens-gone

That's it — it builds nothing on your machine, starts a local HTTP server, and opens your browser at the dashboard. Subsequent runs reuse the on-disk parse caches (~/.cache/tokens-gone/*.json), so warm scans are sub-second even with thousands of session files.

Flags

-p, --port <n>   Port to listen on (default 27821, or $PORT)
--no-open        Don't open the browser
-h, --help       Show help
-v, --version    Print version

Environment

• CLAUDE_CONFIG_DIR — override the Claude data dir (default ~/.claude).
• CODEX_HOME — override the Codex data dir (default ~/.codex).
• XDG_DATA_HOME — where OpenCode keeps opencode/opencode.db (default ~/.local/share) and where Oh My Pi may keep XDG-migrated sessions. Reading OpenCode needs Node 22.5+ for node:sqlite; on older Node versions the other providers still work and OpenCode is skipped.
• PI_CODING_AGENT_DIR — override the Pi agent dir (default ~/.pi/agent).
• PI_CODING_AGENT_SESSION_DIR — override the Pi sessions dir directly.
• OMP_CODING_AGENT_DIR — override the Oh My Pi agent dir (default ~/.omp/agent).
• OMP_CODING_AGENT_SESSION_DIR — override the Oh My Pi sessions dir directly.
• OMP_PROFILE — select an Oh My Pi profile.
• PORT — same as --port.

Why

ccusage works but is slow (tens of seconds on real data) and refetches LiteLLM pricing every run. tokens-gone is the same idea but:

• Cold scan: ~8s for 1.7GB of JSONL (5000+ files); warm: ~700ms.
• Hardcoded Anthropic and OpenAI pricing — no network requests, ever.
• Browser UI with click-to-filter on charts, models, projects, sessions.

How it works

• server/providers/* walk provider-specific JSONL directories, parse logs streamingly, and cache parsed entries on disk keyed by (path, size, mtime). Unchanged files are reused verbatim; appended files are tail-parsed. OpenCode V1 and V2 are the exception: they store sessions in SQLite, so their shared provider reads the DB read-only, deduplicates migrated messages, and caches per session instead of per file.
• server/pricing.ts resolves Anthropic Claude and OpenAI GPT pricing by model name, including provider-prefixed and date-suffixed model IDs. It accounts for provider-specific caching, tiered pricing, and fast-mode rates.
• server/server.ts exposes /api/data (gzipped) and /api/refresh.
• The SPA loads all entries once and re-aggregates client-side on every filter change — fast enough for ~300k entries.

Development

Requires Bun for the dev/test loop. The published package runs on plain Node 20+.

bun install
bun run dev                # backend on :27821 + vite on :5173 (open :5173)
bun run dev:server         # just the backend (--hot)
bun run dev:client         # just vite

bun test
bun run typecheck
bun run build              # vite build + esbuild bundle of the server CLI

License

MIT