Cost Tracking
ClaudeControl tracks token usage and cost for every agent interaction.
Real-Time Cost Display
Header
The total cost of the active workspace appears in the header bar as $X.XX.
Chat Header
Each agent's individual cost appears below its name as $X.XXXX.
Cost Dashboard
Navigate to Settings → Custo e Agentes to see:
- Total USD — accumulated cost across all agents in the workspace
- Total Turns — conversation rounds
- Input Tokens — tokens sent to the model
- Output Tokens — tokens received from the model
- Per-Agent Breakdown — table sorted by cost, showing each agent's contribution
Budget Controls
Max Budget (auto-stop)
Set maxBudgetUsd in runtime settings. When an agent's total cost exceeds this limit, ClaudeControl automatically:
- Interrupts the conversation
- Logs a system message:
"Auto-stop: custo total $X.XX excedeu limite de $Y" - Sets the agent to idle
Max Turns
Set maxTurns to limit conversation rounds per agent session. The SDK enforces this natively.
Cost Alerts
When an agent's accumulated cost crosses a threshold, a system message appears in the chat:
| Threshold | Alert |
|---|---|
| $1.00 | "Alerta: custo total deste agente atingiu $X.XX (>$1)" |
| $5.00 | "Alerta: custo total deste agente atingiu $X.XX (>$5)" |
| $10.00 | "Alerta: custo total deste agente atingiu $X.XX (>$10)" |
Each threshold triggers only once per agent.
Subagent Dedup Detection
When an agent spawns a subagent with a name similar to one already running, a warning appears:
Subagente duplicado detectado: "Task 1" similar a "task1"This helps prevent the model from re-dispatching the same task multiple times.
SDK Auto-Compact
When the SDK detects the context is getting large, it automatically compresses it. ClaudeControl shows:
"Compactando contexto..."— compression started"Contexto compactado (auto, 167235 tokens)"— compression complete
No manual rotation needed — the SDK handles context management natively.
Token Optimization
ClaudeControl optimizes token usage:
| Optimization | Saving |
|---|---|
| Pool context refresh every 10 turns (not every message) | ~450 tokens/turn |
| Cached context files (5-min TTL) | Disk I/O per message |
| Tool results truncated in WebSocket (2KB vs 5KB) | Payload size |
| Log pagination (last 50 entries in API) | Response size |
| Attachments saved to disk (not base64 in logs) | ~170KB per image |