agent-walker

agent-walker

日本語: README.ja.md

Can you see where your AI agents have walked?

A local terminal dashboard built from the logs your AI coding agents already write to your machine — tokens, API-equivalent cost, projects, tools, and autonomy, a month of runs in one screen. Put in the work and you'll earn a codename.

Only the one watching can deter it.

Why

Subscription AI tools don't show you what you actually use. agent-walker reads the session logs on your disk and answers, in one screen:

• How much would this cost on the API? — cache-aware cost windows (today / 7d / 30d / period), so you know if the subscription pays for itself
• Where do the tokens go? — per-repository breakdown across agents
• When and with which model? — GitHub-style activity grass, stacked per-model daily bars, hour-of-day profile in your local timezone
• Can it run unattended? — turn-duration histogram weighted toward the 20-minute-plus autonomy range

Privacy

• Your logs never leave your machine — agent-walker does not phone home with usage or error data.
• The only always-on outbound traffic is a GET of public model-pricing metadata from LiteLLM's pricing database (MIT-licensed) when the dashboard loads or reloads a report. It carries no log content and no usage data, but it does let GitHub's CDN see your IP, the same as any git pull. Block it with a firewall rule or run offline if that matters to you.
• Cursor is the one exception. It's auto-detected when you're signed into Cursor locally, and reading its usage reaches the network: agent-walker uses your local Cursor session token to query Cursor's own usage dashboard (Cursor keeps no usage on disk). That sends your session cookie to Cursor to read your own usage. It's skipped (no request) whenever there's nothing to read — Cursor not installed, or signed out — so nothing is sent unless you're signed in. See docs/cursor.md.
• Release binaries carry GitHub Artifact Attestations:

gh attestation verify agent-walker-aarch64-apple-darwin.tar.xz -R miiiiiiich/agent-walker

Use

No install — npx / bunx fetch the prebuilt binary for your platform (macOS arm64/x64, Linux arm64/x64 GNU, Windows x64) and run it:

npx agent-walker
# or
bunx agent-walker

Key	Action
← → / h l / Tab	switch provider (only those with data; most-used first; Total last)
↑ ↓ / j k / PgUp PgDn	scroll
1–9	jump to tab
r	reload
s	share your stats card (copy image / save to your OS Downloads folder)
q / Esc / Ctrl-C	quit

Flags:

Flag	What it does
--days <N>	Analysis window for the graphs (default 30; the codename's throughput level is always taken from the last 30 days, so the rank doesn't drift with --days)
--share <path>	Render the stats card to a PNG, print its caption, and exit
--no-cache	Rescan every log file, ignoring the parse cache
--no-cursor	Disable the Cursor collector — the only one that sends a credential off the machine (your Cursor session cookie, to cursor.com) — so it makes no network request
--claude-dir / --codex-dir / --agy-dir / --opencode-dir	Point at non-standard log locations (also honors CLAUDE_CONFIG_DIR, CODEX_HOME, and OPENCODE_HOME when set)
--cursor-state-db	Point at a non-standard Cursor state.vscdb (or set CURSOR_TOKEN to supply the session token directly)
--completions <shell>	Print shell completions (e.g. --completions zsh) and exit

Share your codename

Press s in the dashboard (or run agent-walker --share card.png) to render a shareable stats card. Instead of leading with a raw token count, the card reads how you work: a GitHub-style activity grid, your hour-of-day rhythm, and a per-model split, with parallelism and task-time as plain numbers.

Your usage also earns a codename — a rank you climb as you go. Exactly how it's earned is left as a puzzle, and repository names never appear on the card — just glance before you post.

Works with your agents

Every agent is auto-detected — install nothing, configure nothing, just run it.

Agent	Tokens & cost	Models	Projects	Tools	Activity
Claude Code
Codex CLI
OpenCode
Cursor			—	—
Antigravity

Local-only, except Cursor — it keeps no usage on disk, so it's read from Cursor's own dashboard over the network (only when you're signed in; see Privacy). The per-agent pages explain exactly what's read, how tokens/cache/cost are counted, and the caveats.

Everything is parsed in parallel and cached per file (~/.cache/agent-walker/), so warm starts skip the bulk of the work and only reparse log files whose (mtime, size) changed. Log lines are treated as untrusted input — malformed records are counted and skipped, never evaluated.

Keep more than 30 days of history

Claude Code deletes transcripts after 30 days by default, so on first use the dashboard will only show the last month no matter how long the window is. To keep a year of history, add this to ~/.claude/settings.json:

{ "cleanupPeriodDays": 365 }

Cleanup runs on session start, so anything older than 30 days that was already deleted is gone — set this early. Codex CLI keeps its sessions indefinitely; no setting needed.

Notes & limitations

• Cost figures are API-equivalent estimates (what the same usage would cost on metered pricing), not your actual bill. Rates come from LiteLLM with cache reads/writes priced separately; the pricing-fetch date shows up in the COST panel when pricing loads and the terminal has room for it.
• Token totals are cache-inclusive (input + output + cache writes + cache reads). Claude Code re-sends the full context every turn, so for heavy users the bulk of the total tends to be cache reads — real but cheaply billed context re-reads, not new work. See docs/claude.md.
• Antigravity tokens are read from an undocumented protobuf (its per-conversation SQLite), decoded by tag number and checked per row for mutual consistency of the output fields (#3 == #9 + #10), which catches a re-meaning of those output tags; the residual gap is noted in docs/agy.md. They count toward the totals. Cost is estimated from its Gemini model's display name via the same LiteLLM table as every other agent.
• These numbers won't match each agent's own in-app usage display: those use different time windows and count cache reads differently. See the per-agent pages above.
• Antigravity log timestamps carry no timezone and are assumed to be local.
• Windows binaries are published from 0.3 onward; the dashboard runs in Windows Terminal / PowerShell and reads logs from %USERPROFILE%. WSL installs work too — paths just stay in their Linux form.

License

MIT or Apache-2.0, at your option.