agentburn: model tool for Hermes Agent

Local profiler that shows which 5-hour window burned your usage, and where tokens went

What agentburn does

Subscription users hit a rolling usage window, not a flat daily cap, so agentburn surfaces the peak 5-hour window against the median of your own active windows, what filled it by model, by source (you, subagents or scheduled work) and by kind (cache reads, cache writes or output), and measures that peak against your own recorded cut-off moments rather than inventing a threshold Anthropic has not published. Per-token payers, including those running Hermes, instead get a straight cost view: where the money went overnight, which loops or retries drove it, and ready config fixes.

Everything runs locally and read-only against your assistant's own logs, with zero external dependencies and nothing leaving the machine; uvx agentburn runs it directly, and a browser demo requires no install at all. Codex CLI users additionally get the provider's own rate_limits.used_percent reading paired against agentburn's local weighted-usage estimate for the same window.

Key features

Peak-vs-typical 5-hour usage window comparison, not just a running total
Cost-by-source breakdown (direct use, subagents, scheduled jobs) for per-token billing
Measured usage ceiling built from your own recorded cut-off moments, never invented
Zero dependencies, fully local and read-only against existing agent logs
Separate limits and context views for subscription ceilings and long-context cost

When to use it

Finding which 5-hour window actually caused a Claude Code or Hermes rate-limit cutoff
Seeing how much of an overnight bill came from cron jobs versus subagents versus direct chat
Deciding whether a /clear at a given context size would have saved meaningful cost

Who it is for: Anyone running Claude Code, Hermes Agent, OpenClaw or similar agents who wants to know exactly which window or source drove their usage, not just a total.

How it fits with Hermes Agent

agentburn lists Hermes Agent among its supported agents and treats per-token Hermes usage the same way it treats OpenClaw: as a cost problem to localize by source and time, rather than a window-limit problem.

How to install agentburn

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

uvx agentburn

FAQ

What is agentburn?

agentburn is a local, zero-dependency profiler that identifies which specific usage window or source, cron jobs, subagents or direct chat, actually drove an AI agent's cost or rate-limit cutoff.

Does agentburn work with Hermes Agent?

Yes, it lists Hermes Agent among its supported agents and reads its local logs the same normalized way it reads Claude Code, Codex CLI, Gemini CLI, opencode and OpenClaw logs.

How do I run agentburn?

Run uvx agentburn directly for the cost view, uvx agentburn limits for the subscription rolling-window view, or try it with no install at all through the browser demo linked in the README.