Budget
The Budget settings page is where you put a ceiling on cost. Daily and monthly limits, smart routing to cheaper models when appropriate, and alert thresholds so you hear about a spike before it becomes a surprise on next month's invoice.
What this page is for
Exolvra agents call paid LLM APIs. Without limits, a misbehaving agent can burn through hundreds of dollars in an hour — that’s not a hypothetical, it’s the most common first-week incident on any new install. This page is the set of knobs that keeps that from happening.
Three cards: Spending Limits, Smart Model Routing, and Alerts & Thresholds. Save once; changes apply instantly.
Spending limits
| Field | What it controls | Recommended starting value |
|---|---|---|
| Daily Limit ($) | Maximum total spend per calendar day across all agents | $5–$50 depending on workload |
| Monthly Limit ($) | Hard cap for the calendar month | 15–30× the daily limit |
| Auto-downgrade models when approaching limit | When spend crosses the soft threshold, the router silently switches new calls to cheaper models | On by default — leave it on |
Daily limit is a hard stop. When the instance hits the hard-limit percent (see below) of the daily cap, new tool calls and agent turns are refused until midnight UTC. Budget exhaustion isn’t a silent failure — the user sees an explicit message that the cap was reached.
Monthly limit is the upper ceiling. Hitting it refuses all work until the first of the next month. In practice, daily limits catch everything; the monthly cap is belt-and-braces.
Smart model routing
When PM-driven smart model routing is on, the Project Manager agent recommends a model tier for each issue it creates: economy, standard, or premium. Cheap tasks go to cheap models; complex tasks get the premium tier. The recommendation is part of the issue metadata and the assigned agent respects it on pickup.
In practice this saves 40–60% on issue-driven work because most tasks don’t need the frontier. Simple research, documentation, data entry — all fine on economy. Complex reasoning and multi-step codework get the premium model automatically.
Smart routing requires:
- The Project Manager agent to be present (it is by default — the
project-managerspecialist is seeded on first launch) - At least three models configured across tiers (cheap, mid, premium)
With fewer models, there’s nothing to route between. Turn it off if you’re running a single-model deployment.
The alternative to PM-driven smart routing is CommonStack smart routing (see Providers → CommonStack) — that’s router-side, a property of the model string rather than the agent’s choice. The two are independent: you can run both, either, or neither.
Alerts & thresholds
| Field | What it triggers | Default |
|---|---|---|
| High Cost Alert ($) | Single request cost exceeds this — fires an admin alert | $1.00 |
| Session Cost Warning ($) | Cumulative cost of one session crosses this — fires a one-time session alert | $2.00 |
| Soft Limit (%) | At this % of daily budget — auto-downgrade kicks in, admin gets a soft warning | 80 |
| Hard Limit (%) | At this % of daily budget — new work is refused until midnight | 95 |
The soft/hard pair is the main safety valve. At 80% of the daily cap, auto-downgrade engages: the router swaps premium models for cheaper ones on new calls. This gives you a glide path into the wall rather than hitting it full-speed.
At 95%, the wall is real: new tool calls fail with a budget-exhausted error, and the agent sees the error and stops requesting. Existing in-flight requests complete.
High Cost Alert is useful for catching individual runaway calls — an agent that somehow invoked a 1-million-token prompt, a single search that returned 100MB, that kind of thing. If it fires often, something is wrong; investigate the session from Activity Timeline.
Session Cost Warning fires once per session when cumulative session cost crosses the threshold. It’s aimed at the “I left chat open all day” class of accidents. A session that’s genuinely doing expensive work will blow past $2 in an hour — raise the threshold if that’s normal for your workload.
Common pitfalls
Starting without a budget. If you spin up a fresh instance and grant agents full tool access without setting a daily cap, a loop or a badly-prompted agent can rack up real money. Set some limit — even a conservative one — before you leave the instance running.
Setting limits so tight that real work fails. A $1/day cap cuts off after two or three proper research sessions. If agents are failing with budget errors on legitimate work, raise the limit.
Turning off auto-downgrade. It’s tempting — you want to see exactly what you configured. But when you hit the soft threshold, auto-downgrade prevents a hard block; turning it off means you bounce straight from “work fine” to “work refused”. Leave it on.
Ignoring the monthly cap. If the daily cap is $10 and the month has 31 days, the monthly limit should be at least $310. A monthly cap lower than daily × days will bite on the last day of the month when you least expect it.
Where to go next
- Budget dashboard — the day-to-day read of current spend
- AI Models — the allowlist, complementary to budget
- Providers — where CommonStack smart routing lives
- Audit log — where budget-cap events are recorded