How AI requests and tokens are counted

One AI request is one generation run started from the chat panel in Build, Ask, or Fix mode. A run may involve several model calls (plan, edits, repair attempts) but it counts as one request. Cancelled runs still count because the provider has already done work.

After every run the provider reports input and output tokens. Codexa records the sum as an ai_tokens usage event for the workspace. Tokens are counted across all model calls in the run, including repair attempts, so a run that needs two repairs uses more tokens than one that passes validation first time.

Limits apply per workspace, not per user. In a Team workspace all members share the pool.

Before a run starts the server checks the workspace's usage for the current period against the plan (assertQuota). If starting the run would exceed either the request or the token limit, it is refused with a quota_exceeded error that names the limit and the amount used. Usage recorded so far is never adjusted downwards.

Usage periods are calendar months in UTC, starting on the first of the month. Unused requests and tokens do not roll over. Changing plans mid-month applies the new limits immediately to the current period's usage.

Settings, Billing shows requests, tokens, runtime minutes, deployments, and storage for the current period. See usage.