← Back to home

Usage Tracking

What the /usage panel shows, why it varies by provider, and how to read the indicators.

AAA Code surfaces three live indicators next to the composer and a richer panel behind /usage: a context-window meter for the current conversation, plus 5-hour and 7-day subscription windows. What populates depends on which provider you’re using.

The composer chips

  • Context— percentage of the active model’s context window consumed by the most recent turn (input + cache read + cache creation tokens). Auto-promotes to 1M when the SDK reports the context-1m-*beta or when observed usage exceeds the model’s base 200k window. Fills in after the first response in a new session.
  • 5h — five-hour rolling rate-limit window for claude.ai subscriptions. Shows percent used and time until reset.
  • Week — seven-day rolling window. Same shape as 5h.

What each provider exposes

Anthropic via claude.ai login (Pro / Max / Team)

All three indicators populate. The 5h and Week percentages come from the Agent SDK’s rate_limit_event, which fires after each real model turn. If the panel still reads no data yet right after opening, send a prompt — the bars fill in on the first response.

This is the path you get for free if you’ve already run claude loginand don’t configure an explicit provider. See the provider setup guide for the zero-config flow.

Anthropic via API key

Context % populates normally. The 5h / Week chips are hidden and the dialog says the active provider doesn’t expose subscription usage. API key billing is per-token and not subject to the subscription rate-limit windows; check the Anthropic Console usage page for spend.

Anthropic-compatible presets (DeepSeek, Zhipu, Kimi, MiniMax)

Context % populates. Subscription chips hidden — these providers run their own billing dashboards. Check each provider’s console for usage and spend.

AWS Bedrock and Google Vertex

Context % populates. Subscription chips hidden. Bedrock and Vertex bill through their respective cloud consoles; AAA Code can’t see those bills.

Why the AAA Code panel looks thinner than the CLI’s

The terminal CLI’s /usagerenders cost, per-model token breakdowns, and a separate “Sonnet only” weekly window. Those numbers come from a private Anthropic account endpoint that the CLI fetches with your OAuth token directly. The Agent SDK that AAA Code embeds doesn’t expose that endpoint, so AAA Code shows what the SDK gives it: the context window, plus 5h and 7d percentages from rate_limit_event.

For the full ledger, run claude /usage in the terminal or open the Anthropic Console.

Troubleshooting

  • Bars stay empty after several turns. rate_limit_event only fires after a real model round-trip. Slash commands (/usage itself, /clear) don’t count. Send any prompt that hits the model and the bars populate.
  • Context % reads 100% on a fresh session.Usually means a 1M-context Team-tier session was resumed from a JSONL that didn’t carry the context-1m-* beta. AAA Code auto-promotes to 1M on the next turn once it observes usage above 200k.
  • 5h shows 0% with a reset time, but you know you’ve been busy. The SDK reports a status event without a numeric utilization on fresh windows; AAA Code renders it as 0% rather than blank. The number updates as soon as a real rate-limit event includes utilization.