agent-walker

日本語: README.ja.md

Want to know how you use AI, not just how many tokens you burn?

agent-walker reads the logs Claude Code, Codex CLI and friends already leave on your machine, works out every metric it can, and puts them on one screen. Compare with your friends, share it, have fun! #agent-walker

bunx agent-walker
# npx agent-walker

What you learn

More than each agent's status or usage command tells you.

Every section reads the same window: the last 30 days.

Question	Section
Which days you use it most, and at what hours	ACTIVITY / BY HOUR
How the daily burn splits across models	TOKENS PER DAY
Which projects, models and settings eat the tokens	PROJECTS / MODELS / MODES
Which tools and subagents do the work	TOOLS / SUBAGENTS
Which skills the tokens went to (Claude)	SKILLS
How many agents you run at once	PARALLEL AGENTS
How long you actually let an agent work	TURN LENGTH
How many hours the agent worked, and how fast you answer it	WORKING TIME
Whether you're wasting context	CONTEXT
What it would have cost on the API	COST
How close you got to the plan limit (Codex)	LIMITS
How many AI credits you spent (Copilot)	CREDITS
Your favourite model, best day, peak hour, longest session, streak	SIGNAL

Share

Press s and a shareable image lands on your clipboard. Nothing private on it, no project names.

Your rank and animal change with token volume. The colour changes with the hours you work.

Supported agents

Claude Code / Codex CLI / OpenCode / Cursor / GitHub Copilot CLI / Grok Build / Antigravity. All auto-detected.

Each one has a page on what is read and how: Claude Code / Codex / OpenCode / Cursor / Copilot / Grok / Antigravity

Privacy

agent-walker sends your logs nowhere. The only network use is fetching the price table (kept for the day once it lands), and fetching Cursor's usage when you're signed in.

Keys

Key	Action
digits	switch tab
r	reload
s	copy share image
q / Esc / Ctrl-C	quit

Flags

Flag	What it does
--share	write the share image to a PNG and exit
--json	export the summary and dated events as JSON (experimental)
--days <N> (with --json only)	window in days (default 30); TUI stays at 30
--no-cache	reread every log

JSON

--json writes the same numbers as the TUI, plus the dated events behind them, as one JSON document. Experimental: the schema may change between minor releases.

agent-walker --json | jq '{schema_version, window, total}'

Note

Costs are API-equivalent estimates, priced with LiteLLM's table.

Acknowledgements

ccusage pioneered local usage tracking for AI coding agents and first identified and solved problems that agent-walker also addresses. See THIRD_PARTY_NOTICES.

License

MIT or Apache-2.0, at your option.