Token Budget

Reeboot tracks token usage and LLM cost for every agent turn, and can enforce limits at three granularities: daily, per-session, and per-turn. All limits are per-context — each context has its own independent budget bucket.

Budget Layers

Set null (the default) to disable a limit.

Configuration

{
  "budget": {
    "daily_cost_usd": 2.00,
    "session_tokens": 100000,
    "turn_tokens": 20000,
    "warn_threshold": 0.8
  }
}

When warn_threshold is reached (default 0.8 = 80% of limit), the agent is notified. At the limit, the turn is blocked and the agent is instructed to wrap up.

Cost Tracking

Cost is calculated using pi’s built-in ModelRegistry, which includes per-token pricing for Anthropic, OpenAI, Google, Groq, and other major providers.

For local models (Ollama) and providers without pricing data, cost is shown as “cost unavailable” — not $0.00 — to avoid misleading reporting.

Agent Tools

check_budget

For the agent itself — returns the status of the active per-task budget (set via set_budget):

check_budget()
→ { spent_tokens: 1234, limit_tokens: 5000, remaining: 3766 }

Returns “No active task budget” if no task budget is set.

set_budget

The agent sets its own per-task budget for the current session:

set_budget(tokens: 5000, description: "research task")

Resets at the end of the session (agent_end).

budget_status

For the owner — queries historical spend:

budget_status({ period: "today", operationType: "user_message" })
→ "Today: 4,231 tokens / $0.18 (user messages)"

budget_status({ period: "this_week" })
→ "This week: 42,800 tokens / $1.84 (all operations)"

Configuration Reference