agentburn / playground

A simulation with sample data — click a command or type it. The real tool runs locally and reads your agent's own logs; nothing here is your data.

~ agentburn playground
$
try the real one: uvx agentburn  · zero deps · local-only · read-only

Supports Hermes Agent, OpenClaw and Claude Code · GitHub · token-history (the world side)

FAQ

How do I see where my AI agent spends tokens and money?
Run uvx agentburn. It reads your agent's own local logs — Hermes Agent (~/.hermes/state.db), OpenClaw (sessions), Claude Code (transcripts) — and breaks spend down by source: cron, telegram/discord gateways, cli, subagents, plus a night-burn strip (00:00–08:00). Local-only, read-only, zero dependencies; nothing leaves your machine.

Why is my agent's bill so high at night?
Usually scheduled work: cron jobs and heartbeats on an expensive model while you sleep. agentburn fix proposes paste-ready config patches (cheap model for cron, heartbeat activeHours) with estimated $/month savings — dry-run, it never edits files itself.

What is the agent actually doing with all those tokens?
agentburn why reconstructs behavior from the agent's records: functions called, re-read loops (same file paid N times), retry storms, idle heartbeats, per-cron-run receipts, context-window compactions. Observations with evidence, not verdicts.

How is this different from ccusage or built-in /usage?
Those total one agent's spend. agentburn is cross-agent (Hermes, OpenClaw, Claude Code), attributes spend to sources and behavior, compares your models against world usage trends (agentburn drift), and ranks your efficiency against anonymous community percentiles (agentburn rank).