Every figure here is notional — what these tokens would have cost at published API rates. On a Max or Antigravity subscription that is not money: the bill is flat however much you use, and a local model is free. It is still the only number that puts a subscription, a metered key and a local model on the same scale.

Prepaid balances on external services (ElevenLabs, Suno, Meshy), polled every few minutes. Real money, unlike the notional token costs. Spend is estimated from the balance falling between polls — the pools are used outside the gateway too — priced at a cost per credit you can edit. A low balance raises a Telegram alert.

The TUI engine runs each job in a real Claude Code terminal held open by the tui-host (supervised by this gateway). Set an account's engine to claude-tui on the Accounts tab to route its jobs here; a job whose host is down still falls over to the SDK. Each account needs its interactive login once — driven from here, no terminal.

Every run, grouped by session. Click a session to see its jobs; click a job to see exactly what went in (prompt + options) and what came out — the full event trail: tool calls, results, errors, tokens, cost.