Documentation Limits & quotas

Limits & quotas

Agent runs are charged by actual AI usage from your prepaid credit balance, which is shown in the chat header and refreshes after each response. A run costs what its model calls used — input, output and cached tokens across every tool-loop step — so larger models (Claude Opus) cost more than smaller ones (Claude Haiku, Gemini Flash), and a short question costs a fraction of a credit. Fractions accumulate and are deducted as whole credits. Runs on your own Anthropic API key (Settings → API Keys & Integrations) are free. New accounts get 25 free credits. When the balance runs out, AI requests are refused with an "out of credits" message until you top up. Each run is also bounded by the limits below.

Limit
Value
Tool-loop iterations per message
200
Response time per message
2 minutes
Gap with no streamed output
30 seconds
Request size
4 MB
Attached file context
8,000 characters per file, 100,000 in total
Files read in one parallel read
20
Matches per search
50
Large-file warning
Over 500 lines (warning only, the write still happens)
Versions kept per file
50
Note
Credits don't expire. Check your balance and buy more in Settings → Usage & Billing.
What uses credits (charged by usage)
  • Messages to the Claude Code Agent and the AI chat — by the tokens the model reads and writes
  • Automated CI fixes and Visual Edit prompts — by tokens
  • Voice replies (text-to-speech) — by characters
What doesn't count
  • AI requests that run on your own Anthropic API key
  • File saves, terminal commands, and manual code edits
  • Direct text, color and font edits in Visual Edit
  • Deployments, previews and browsing
Safety behaviour
  • write_file blocks content that looks like an API key, token, password or private key. Other edit tools don't scan automatically.
  • File paths containing "..", absolute paths and backslashes are rejected.
  • If the agent calls the same tool with the same arguments 3 times within its last 6 tool calls, the call is rejected and a loop warning is shown.
  • Destructive-action warnings notify you but don't pause the run — press stop if you don't want the agent to continue.
  • The agent is instructed never to claim that tests or builds passed, because it cannot run them.