Your Claude limit, explained. Free. One command.
Burn is a free tool for Claude Code and Codex that runs on your computer.
npx agentguard-burn whyPaste this into your terminal on the computer you use for Claude Code or Codex.
Burn keeps usage numbers on your computer. An optional anonymous install count sends no usage. What it sends.
See it work
Screens from Burn and the free AgentGuard plugin. Each one says what it is and when it was captured.
Status line · real session · Oct 7, 2026
Above your prompt, live: sub-agents, tokens, the share that went to sub-agents, and your 5-hour limit.
This line comes from the free AgentGuard plugin for Claude Code, which runs on Burn. Install the plugin.
Video · Oct 7, 2026
The line at work: its numbers are from a real session, and the climb to 40 sub-agents is simulated.
/agentguard pane · preview build, sample data · Oct 1, 2026
Type /agentguard for the detail: both sub-agent windows, tokens, your plan's reset, and each launch, ask and stop as a signed row.
18509 assistant responses | 7,610,368,479 tokens
Bucket Tokens Session tokens Last turn tokens
Instruction stack + system 3,050,957,858 40.1% 66.9%
History re-sent 546,976,085 7.2% 33.0%
Re-read files 185,331 0.0% 0.0%
Subagent fan-out 3,904,343,726 51.3% 0.0%
Tool output 6,971,816 0.1% 0.0%
Conversation 3,767,644 0.0% 0.0%
Full-prefix rewrites 77,229,519 1.0% 0.0%
Output 10,805,641 0.1% 0.0%
Unattributed 9,130,859 0.1% 0.0%
Sub-agents that ended without finishing: 10 of 154, 194,735,481 tokens (2.6% of session tokens)
Burn why report · our long session · Oct 3, 2026
What the one command prints: every token in the session, split by cause.
Cropped from the output of npx agentguard-burn why, Burn 0.3.27, captured Oct 3, 2026. The session ID, the usage line, the List USD column and the dollar totals are removed; nothing else is changed. Token shares partition recorded usage; byte sizes only split measured token increments and are never converted into tokens. Method in the docs.
Where this session's tokens went
One long Claude Code session, captured Oct 3, 2026 with Burn 0.3.27. These rows come from its report.
- 7.6 billion
- tokens recorded in the session (7,610,368,479).
- 51.3%
- went to the sub-agents Claude started.
- 40.1%
- went to the setup Claude re-reads every turn: the system prompt, tools and CLAUDE.md.
- 10 of 154
- sub-agents ended without finishing. They still used 194.7 million tokens.
What to check next
Look at which sub-agents ended without finishing, and what each one was started to do. In this session, 10 of 154 did, and they still used 194.7 million tokens.
Run it on your computer
Paste this into your terminal on the computer you use for Claude Code or Codex. It covers your last session.
Run Burn on the computer you use for Claude Code or Codex. It covers your last session.
npx agentguard-burn whyOther commands
Usage across sessions
Every Claude Code and Codex session saved on this computer.
npx agentguard-burn calendarMake a shareable image
An image of the calendar totals that you can share.
npx agentguard-burn cardWhat the numbers mean
Not your allowance. Burn reports the tokens a session recorded. Shares of those tokens are not a reading of your plan allowance, and they are not savings.
One computer. Burn reads the sessions saved on the computer it runs on. Other devices are not included.
Two tools. Burn explains a session from the records on your computer. The free AgentGuard plugin runs on Burn and shows the live line above your prompt.
Usage stays local. Burn keeps usage numbers on your computer. An optional anonymous install count sends no usage. What it sends.
npx agentguard-burn why