Skip to main content
Every LLM call is proxied through the mesh; usage is recorded in a SQLite WAL cost ledger at data/costs.db. Costs are computed via estimate_cost() against LiteLLM’s model registry, then enforced against per-agent budgets.

​
How It Works

  1. Agent makes an LLM call through the mesh proxy.
  2. The mesh checks the agent’s remaining budget before forwarding to the provider.
  3. If the agent has exceeded its daily or monthly budget, the call is rejected.
  4. After a successful call, token usage and dollar cost are recorded in data/costs.db.
Because agent containers hold no provider keys and all calls proxy through the mesh, an agent cannot bypass the budget check.

​
Configuring Budgets

Set per-agent budgets in config/agents.yaml: 
agents:
  researcher:
    budget:
      daily_usd: 5.00
      monthly_usd: 100.00
Defaults: 10/day,10/day, 200/month per agent. Allowed range: daily 0.01–0.01–1000, monthly 0.10–0.10–30000. Values outside this range are rejected at config-load time. When an agent exceeds its daily or monthly budget, subsequent LLM calls are rejected until the budget resets.

​
Per-project rollups

Costs are tracked per-agent and per-project. The cost ledger keys on both, so you can see what each project is spending in aggregate across all its agents. Per-project rollups surface in the dashboard at Settings → Costs.

​
Viewing Costs

From the interactive REPL: 
/costs
Shows per-agent spend for today, this week, and this month, including token counts and dollar amounts. In the dashboard, Settings → Costs shows per-agent and per-project breakdowns with a period selector and per-agent budget bars.

​
CAPTCHA Costs

CAPTCHA solver spend is tracked separately from LLM cost, in millicents (1/100,000 USD), in data/captcha_costs.json (chmod 0o600). Both per-agent and per-tenant monthly USD caps apply, with alerts at 50% / 80% / 100% of the cap. The fleet-wide kill switch CAPTCHA_DISABLED stops all CAPTCHA solving regardless of budget.

​
Model Failover and Costs

When using failover chains, cost tracking follows the call to whichever model actually handles it. If the primary model fails and the request cascades to a cheaper fallback, only the fallback cost is recorded against the agent’s budget.