Skip to content
All Docs pages

Reference

Usage and cost

A local ledger of one row per turn, so Fleet can answer what you spent on which model, in which repo, on what.


Fleet bills you for nothing. Model usage is billed to your own Claude account, through your own Claude Code installation. What Fleet does is keep a record so you can see it.

Two different numbers

A session carries a running total: what this session has cost so far. That answers "is this session expensive" and nothing else, because it keeps no history and no breakdown.

The usage ledger is separate and can answer the rest. It is append-only, one row per completed turn, holding:

  • when
  • which session and which repository
  • which model served the turn
  • cost in dollars
  • input, output, cache read and cache write tokens
  • whether the session was solo, an orchestrator, or a worker

The role field is there so a squad's true cost can be separated from solo work, which is the question you will actually want answered after your first squad.

Cost is recorded per turn, never as the running total. A ledger row that copied the session total would bill every turn for all of its predecessors, and every chart would double count.

What you can ask it

Group by day, by model, by repository, or by session, over any window. Days are days in your own timezone rather than UTC, because a chart of your work should agree with your calendar.

Labels are resolved for display, so you see a repository name and a session title rather than opaque identifiers.

The spend warning

You can set an advisory threshold: Fleet warns once the last 30 days pass it. Zero turns it off.

It is advisory only. Nothing is ever blocked, and Fleet has no way to block it in any case, because the spending happens through your Claude account and not through Fleet.

Where it lives

On your machine, with everything else. The ledger is never transmitted anywhere. Fleet has no telemetry and no analytics.

Keeping the bill down

  • Idle agents cost nothing. A squad between rounds is cheap: an agent that requested review is stopped, not polling.
  • Effort is a per-session setting, from low to max where the model supports it.
  • The agent cap per squad counts the round, so a long-lived squad does not accumulate spawns against it.
  • One agent is often the right answer. See one agent or a squad.