Costs and usage

Every run reports its own cost breakdown to Mission Control while it executes: live per-model samples during Claude runs, and a final per-execution entry when each agent execution ends. That normalized reporting drives the cost figures on a run’s page and needs no setup.

Separately, the host keeps a durable, ccusage-readable archive of the Claude usage its workers generate, so ccusage on the desktop machine can account for remote runs it would otherwise never see.

The host usage archive

Claude executions run on Linux workers, and each session’s transcript lives inside that run’s sandbox home on the worker. After every agent execution (including failed, cancelled, usage-limited, retried, and follow-up executions, and again when an attached terminal exits), the worker exports a minimized snapshot of the session and Mission Control archives it at:

~/.brevi/ccusage/claude/projects/<project-key>/<session-id>.jsonl
~/.brevi/ccusage/claude/projects/<project-key>/<session-id>/subagents/agent-<hash>.jsonl

The snapshot keeps only what usage accounting needs: event type and timestamp, session, request, and message ids, the model, the four token counts, and any pre-calculated cost. Prompt and response text, thinking, tool calls and results, repository content, credentials, and filesystem paths never leave the worker. A re-exported session (a retry, a resumed terminal) replaces its file wholesale, so nothing is ever double-counted. Sessions that used subagents export one extra small file per subagent transcript, nested under that session’s subagents/ directory, so the archive’s totals cover the whole run, all models.

Querying with ccusage

ccusage reads custom data roots through CLAUDE_CONFIG_DIR. Only Brevi worker usage:

CLAUDE_CONFIG_DIR="$HOME/.brevi/ccusage/claude" \
  ccusage claude daily

Local Claude Code usage and Brevi worker usage together, via comma-separated roots:

CLAUDE_CONFIG_DIR="$HOME/.claude,$HOME/.config/claude,$HOME/.brevi/ccusage/claude" \
  ccusage claude daily

Any other ccusage command (monthly, session, blocks) works against the same roots.

Retention

The archive is accounting, not run state: it lives outside ~/.claude, ~/.brevi/runs, and ~/.brevi/workspaces, and no run cleanup (artifact deletion, sandbox reaping, archiving a run in the dashboard) ever touches it. It grows by one small file per Claude session, plus one per subagent transcript that session used, and is kept indefinitely. Reclaiming the space is an explicit operator action: delete ~/.brevi/ccusage, which only forgets historical usage and affects nothing else.