Run costs

The panel can show what each run actually cost — tokens in/out and an approximate dollar amount — right next to the run, so you can answer "how much did last night's run cost?" and "which repos burn the most?" without leaving the dashboard.

Where you see it

Run list — a compact ~$0.42 chip on each row that has cost data.
Run drawer — a Cost line with the full summary, e.g. ~$0.42 · 20k in · 2.8k out · 2 models.

Amounts are prefixed with ~ because they're best-effort estimates, not an invoice. Sub‑dollar runs show four decimals (~$0.0594); larger ones round to cents. Token counts are abbreviated (980, 15k, 2.4M).

How it's measured

Cost capture runs entirely on your worker via tokscale — a tool that reads the local usage data your AI harness already writes (Claude Code, OpenCode, Codex, etc.).

For each run, the worker snapshots cumulative usage at start and at end and reports the difference — that delta is the run's token/cost, sent to the panel and stored on the run. Nothing is estimated from prompt text; it's the harness's own accounting.

BYOK-safe. tokscale reads token usage counters only. It never reads, logs, or transmits your provider keys — consistent with BYOK. Cost data is the only thing that leaves the worker.

Enable it on a worker

Cost capture is opt‑in per worker and fully optional — a worker without tokscale runs exactly as before, just with no cost line. To turn it on, install tokscale on the worker host so it's on the worker's PATH:

# install (Bun) — any install that puts `tokscale` on PATH works
bun add -g tokscale

# confirm the worker can see it
tokscale --version

The next run on that worker will start reporting cost. No panel setting and no worker restart are required — the worker detects tokscale at run time.

Caveats

No tokscale → no cost. Runs are unaffected; the Cost line is simply hidden.
Concurrent runs share the delta. Because tokscale usage is machine‑global, two runs executing at the same time on one worker each over‑count by the other's usage. AM's per‑repo runs are typically sequential, so this is close enough — and it's advisory, never used to block or bill.
Estimate, not billing. Your provider's actual charge can differ (negotiated rates, volume discounts). Treat the figure as a relative signal.

Aggregate & export — the Cost tab

Per‑run cost is useful, but to answer "what did AM cost us this month?" or "which repo (or which profile) burns the most?", open the Cost tab.

Date range — Today / 7d / 30d / 90d, or a custom from/to.
Summary — total cost, run count, and average cost per run for the range.
By repository and by profile preset — so a team can see, e.g., how much the unit-tests profile cost on a given repo.
By model — a simple bar breakdown of spend per model.
Export CSV — one row per run (run_id, repo, started_at, status, cost_usd, preset, model) to reconcile against your provider bill.

Behind it: GET /api/runs/cost-summary and GET /api/runs/cost-export (CSV or ?format=json), both accepting from / to / repo_id.