Usage

A local, always-on record of every LLM and agent call — tokens, cost, and request volume across chat, the Cairn Agent, subagents, automations, and the one-shot AI features. Nothing leaves your machine.

Opening Usage

Click Usage in the sidebar (it sits with the views, or in the Workspace section at the bottom). The view is workspace-level (not per-project) and records everything locally, on every call.

What's tracked

Every request that hits a model is recorded with its input, output, and reasoning tokens, the model id, your saved provider's name, and cost where available:

• Chat — messages in the AI chat panel
• Agent — the Cairn Agent (pi-agent) loop
• Chat subagent / Agent subagent — dispatched research and write agents
• Automation — scheduled automation runs
• One-shot AI — PRDs, commit messages, PR descriptions, explain-code, Idea Flow summaries, compaction, and the AI Tool Builder

The Writing Style wizard is attributed to its own "Writing style" source. Use the source filter dropdown in the toolbar to narrow to any single surface. The layout adapts to the panel's own width, so toolbar controls wrap and stat cards reflow in a narrow pane.

Toolbar & ranges

• Stat cards — input tokens, output tokens, cached input (with % of input), total cost, and request count for the selected range, each with a delta vs. the previous period.
• Chart — daily totals with a Tokens / Cost / Requests toggle. The token view overlays input and output, with a hover breakdown.
• By model — where spend is going, with each model's provider logo, proportional bar, and cost.
• History table — requests are grouped by thread: each chat thread, agent session or automation run is one expandable row showing its title, request count, models, total tokens and total cost. Click a row to see the individual requests underneath (exact time, model with logo, input / output / reasoning tokens, provider, source and cost). One-shot calls with no thread (commit messages, explain, etc.) still appear as single rows. Thread totals follow the range, source and estimates filters.
• Range — 7D / 30D / 90D / All.
• Estimates toggle — on by default. Hides only the estimated dollar figures (shown as —); all calls and token stats stay visible.

How cost is shown

Cost is shown whenever the provider reports it (e.g. NeuralWatt's energy accounting, or credit balances recovered from your account), and otherwise estimated from models.dev pricing — estimated figures are marked with a ~ and can be filtered out with the Estimates toggle. Models whose ids carry region, date or reasoning suffixes (such as -gcp, -0731 or :thinking) are matched to the underlying model's pricing, so they get an estimate too. Estimates are cache-aware: cached input is priced at the model's cache-read/write rates rather than full input, so cache hits aren't over-billed. Costs are always shown in USD. The total reflects only provider-reported cost; estimated totals are not mixed in.

Prompt cache

The Usage view tracks prompt caching across every surface. Input counts new, cached and cache-write tokens, and cache writes are recorded and priced. A Cached input stat shows cached tokens and the percentage of input that hit the cache, and the history table has a Cached column per call. Cache-hit percentages are colour-coded — grey (no cache) → amber → accent → green as the rate rises (≥50% is green) — so you can see at a glance how much of your prompt spend caching is saving.

Clearing recorded usage

Use the trash button in the toolbar to wipe the token/cost/request history for the current workspace. It asks for a second click to confirm (styled tooltips explain each toolbar action) — useful for a fresh month or before handing a device over. The underlying database also auto-vacuums, so heavy use never leaves the file stuck at a large high-water mark.