Skip to main content
Every agent has two server-enforced limits: a daily spend cap (soft pause) and an hourly action cap (hard pause). Both are configured per-agent. Neither can be exceeded — the runtime checks before each LLM call and before each mutating tool call.

​
Daily spend cap

The runtime maintains agent_cost_ledger per (agent, day, provider, model). Before each LLM call, the CostGovernor computes estimated_cost = prompt_tokens × 1.2 × per-token price and refuses if ledger_today + estimated > cap. Refusal sets status=PAUSED_COST, files a COST_PAUSE pending item, and exits the cycle.

​
Hourly action cap

An “action” is any mutating tool call (everything in the Mutating category in Rules and decisions). Read-only tools don’t count. The cap exists to prevent runaway loops — a misbehaving prompt that retries the same action 100 times gets stopped at 60.

​
80% warning

When daily spend crosses 80% of the cap, the agent fires a on_cost_threshold notification. The UI shows an “approaching cap” banner on the agent header. Configurable via the rule set:

​
Model choice

The fast path is opt-in via the rule set:
Default true. Saves ~10× on the common (event-driven) path.

​
BYO key

Bring your own Anthropic or OpenAI API key. We never bill you for usage on the BYO path — your account is charged by the provider directly. See the full walkthrough: BYO Anthropic key.

​
Pro Max downgrade

If your workspace downgrades off Pro Max, all ACTIVE agents move to PAUSED_USER with reason "Plan downgraded — Pro Max required". The agent rows stay in the database — re-upgrade reactivates them with config intact, including rule sets, channels, data sources, and pending items. Memory notes and run history persist throughout.

​
Cost dashboard

Two views in the app:
  • /workspace/agent-costs — workspace-wide MTD total, sum of caps, projection.
  • /agents/{id}/costs — per-agent breakdown by day and model.
Cost dashboard

​
API