AgentGuard Burn: what ate my limit

AgentGuard Burn is a free command line tool that reads Claude Code and Codex session transcripts on the user's machine and shows where the tokens went. It runs locally. Nothing is uploaded.

When to use

• The user hit their Claude Code or Codex usage limit and wants to know why.
• The user asks where a session's tokens went, or why a session was so expensive.
• The user suspects sub-agents ran away.

Step 1: run the report

Run this in a shell on the user's machine:

npx agentguard-burn why

• With no argument it reads the most recently modified local transcript, Claude Code or Codex. Sub-agent transcripts of a Claude Code session are included.
• To pick one session, pass its id, the end of its id, or its transcript path: npx agentguard-burn why SESSION_ID.
• Inside Claude Code the user can run it themselves with ! npx agentguard-burn why.
• The first run downloads the package from npm.

Step 2: explain the report

The report has three parts. Read them to the user in plain words.

1. Header. The session, the number of responses, total tokens and a dollar figure labeled "API list equivalent".
2. Table. Where the tokens went: instruction stack and system (setup read before every message), history re-sent (the conversation re-read on every message), re-read files, sub-agent fan-out, tool output, conversation, full-prefix rewrites (cache reloads), output and unattributed. Point out the one or two biggest rows by token share.
3. The last lines. A Measured change: line may appear after a /compact, a model switch or a /clear. It states a measurement, not a saving. The report then ends with at most one line that starts with Next:. That line is the one action that matters most for this session, chosen from its own numbers. Quote it to the user as written. If there is no Next: line, nothing stood out, so say that instead of inventing advice.

Rules when you explain it:

• Dollar figures are API list price, not a bill. Say "API list equivalent". A user on a Claude or ChatGPT subscription is not charged these amounts; for them, the token columns are what count against the limit.
• Token shares are shares of tokens, not of dollars.
• Never promise a saving. Burn states measurements and one next step.
• A sub-agent that "ended without finishing" was recorded as failed, killed or stopped. Burn does not judge whether any work was right.

Step 3 (optional): the free AgentGuard plugin

Offer this only when the Next: line points at sub-agents (it mentions AgentGuard's free plugin and agentguard.run/try), or when the user asks for a limit on sub-agents. Ask before installing anything.

Claude Code:

claude plugin marketplace add MerchantGuard/agentguard-codex-plugin
claude plugin install agentguard@agentguard

Codex and the ChatGPT desktop app (local tasks):

codex plugin marketplace add MerchantGuard/agentguard-codex-plugin
codex plugin add agentguard@agentguard

Then run npm ci in the installed plugin folder that Codex reports.

What the plugin does, so you can describe it accurately:

• Shipped limits: 15 sub-agents in any 15 active minutes, 40 in any 120 active minutes, 5 billion tokens in one session, and sub-agents at most two levels deep.
• Claude Code in default, accept edits, auto or plan mode: the next launch past a limit waits behind Claude Code's own permission prompt for the user's yes.
• Codex, and Claude Code in bypass permissions, don't ask or print mode: the launch is refused (blocked) with a STOP box. Codex does not ask the user; it blocks until the user allows it.
• To continue past a STOP, the user runs the override command the STOP box prints, in their own shell (with ! in front inside Claude Code, or in a terminal for Codex). You cannot run it for them: the plugin refuses an agent command that tries.
• npx agentguard-burn shadow switches the machine to record-only; npx agentguard-burn enforce switches it back.
• It is free on one machine. Its hooks run locally and open no network connection. Free use sends nothing unless the user asks for an AgentGuard Score.

Other Burn commands

• npx agentguard-burn calendar: tokens per day across Claude Code and Codex, with sub-agent days marked.
• npx agentguard-burn compare: two sessions side by side.
• npx agentguard-burn rewrites: every full-prefix cache rewrite with its cause.
• npx agentguard-burn quiet on: hide the Next: tip; quiet off shows it again.

Do not

• Do not call a dollar figure a bill, a charge or a cost the user paid.
• Do not say Codex asks the user at a limit. It blocks.
• Do not run Burn's override, shadow or calibrate commands on the user's behalf.
• Do not install the plugin without the user's yes.
• Do not copy transcript contents anywhere. Burn reads them locally; keep it that way.

More: https://agentguard.run/burn