Token Garden: where your AI tokens go

Coding agents spend tokens quickly, and the numbers live in log files nobody reads. Token Garden reads the local session logs from Claude Code, Codex, and OpenCode and turns them into one dashboard: tokens over time, estimated cost, cache hit rate, the most expensive sessions, and when in the week the work actually happens.

Token Garden dashboard in dark mode: three gnomes showing live token rates per tool, a grid of usage and cost tiles, and token and cost charts over the last six hours. — The dashboard over a six-hour window. Each gnome is one tool, working at the speed of its live token rate.

Nothing leaves your machine

Session logs contain prompts, file paths, and project names, so the dashboard is deliberately local. It is a small Node.js server bound to127.0.0.1, reading logs read-only. It rejects requests from other websites, and its content security policy stops page scripts from sending data anywhere else. The only outbound request is an hourly download of public model prices, which can be switched off.

No dependencies

There is no npm install and no build step. Clone it, runnpm run usage:serve with Node.js 22 or newer, and open the local URL. The first run parses every session log; after that a per-file cache means a restart takes about a second and only changed files are re-read.

Counting honestly

Most of the token volume in agent work is cache reads, which cost a fraction of fresh input. A single "total tokens" number hides that, so Token Garden also shows active tokens — fresh input plus output — alongside cache hit rate and cost coverage. Costs for Claude and Codex are estimates from public price tables and are labelled as such; OpenCode records its own.

The gnomes

The quirky part is on purpose. A crew of garden gnomes, one per tool, swing their tools faster as tokens flow and fall asleep when a tool goes idle. Smaller characters run the tiles: a hamster wheel for burn rate, an accountant stamping cost coverage, a weightlifter whose barbell grows with session size. A gnome wanders the page. All of it respects reduced-motion settings.

Crews are plain script files with a documented contract, so you can ask your own coding agent to draw a new one and keep it private.

Built to be extended by AI

The repository is small and written to be read by coding agents as much as by people. It includes an agent guide, a harness adapter guide for adding tools such as Cursor, and a crew guide. The README's starting prompt is short:

Clone https://github.com/arturochang/token-garden, read README.md andAGENTS.md, verify Node.js 22+, then start Token Garden with npm run usage:serve and tell me the local URL.

Token Garden is free and MIT licensed. Like everything in LAB256's tools section, it comes without support or a roadmap.

View on GitHub