Where it shows up
Each session's meter — ⚡ 12.4k · $0.42 — appears on its sidebar row next to the model name, and in a status bar under the terminal panes. Open the session info panel for the full breakdown: input and output tokens, cache reads and writes, cost, model, permission mode, last-turn latency with a sparkline of recent turns, plus running counters like active time, tool calls, lines changed, commits and PRs. All of it is mirrored to the iPhone too.
Setup per agent
- Claude Code — automatic. Gallager configures OpenTelemetry in the sessions it launches; you get tokens, cost, latency and model with zero setup. The telemetry is content-free — usage numbers, never your prompts.
- Codex CLI — on by default, with an "Export telemetry" opt-out in Settings → Agents. Codex reports tokens and latency but no cost, so the meter shows tokens alone. Your
~/.codex/config.tomlis never touched. - opencode, pi and omp — included with their plugins; installing the plugin is the whole setup.
How it counts
The meter is per session, not per pane: it follows whatever session is live in the pane and starts fresh when that changes — /clear, a resume, or a new session. The headline number is input + output tokens and deliberately excludes cache reads and writes (agents re-report the whole cached context every turn, which would inflate the number several-fold); the full cache figures are in the info panel. Costs under a cent display as<$0.01.
Good to know
- Everything flows over a local loopback receiver on your Mac — telemetry never leaves the machine except encrypted to your own paired devices.
- Only sessions Gallager launches are instrumented; a
claudeyou start in some other terminal won't report here. - The live meter resets with its session. Longer-horizon per-project usage is aggregated separately and survives restarts.