Token cost and budget for your AI coding sessions

See what your current Claude Code session has cost so far:

/cost:report

What it does

/cost:report reads the token usage of the active Claude Code session, prices it per model, appends a record to agents/cost-tracking/sessions.jsonl in your repository and checks it against your budget. It counts a turn Claude Code replays into several transcripts once, and it splits the main session from subagents.
Budgets are daily, weekly and monthly ceilings in US dollars, set under cost.budgets in your settings. The report warns as you approach a ceiling.
cost.enforcement decides what happens at the ceiling: advisory (the default) shows the breach and keeps working; hard-stop refuses to start the next turn until the budget is raised or reset.
/cost:profile changes how much Agent Config loads on every task — the biggest lever on what each task costs.

How much Agent Config itself costs

Checking its work costs tokens. Measured on Claude Haiku only: the default setting costs 3.3× a run with no rules on the task family where it helps and about 1.7× over the full corpus; loading everything costs 11.7×. On a strong model the gain was a null. Benchmark & cost, explained has the table and its limits.

When to use it

At the end of a long session, before you start the next one.
When you set up a team budget: put the ceilings in the shared settings and pick hard-stop if a breach must stop work.

Example

cost:
  budgets:
    daily: 20
    monthly: 300
  enforcement: advisory

What it needs

/cost:report reads Claude Code’s session files; on other tools it reports that no Claude Code session was found.
Nothing leaves your machine: the records stay in agents/cost-tracking/.

Limits

A model the price table does not know is costed at zero and flagged, so the total is understated until you supply its rate.
Budgets cover what the report has recorded, not your provider’s invoice.