The first question anyone asks before letting an AI agent run on its own is not "can it build something useful?" It is "what happens to my bill if it goes wrong?" Autonomy is exciting right up until you imagine an agent stuck in a loop, spending money while you sleep.
That fear is the real barrier to adopting agent-built software. So it is worth explaining, plainly, how the money works: what a credit is, why a budget cap matters more than a low price, and how to let an agent build freely without ever risking a surprise.
Quick answer
Agent compute is billed as usage credits — a monthly budget, measured in dollars, that the agent draws down as it works. Every plan also has a budget cap, so when the credits are spent the agent is stopped rather than allowed to keep spending. On the free plan the cap is a hard stop with no bill possible; on paid plans you can allow billable overage or keep a strict ceiling. The result is a system where an agent can build autonomously and still never produce a surprise invoice.
Key takeaways
- A credit is a dollar of usage budget the agent spends as it builds and runs software.
- A budget cap is the safety feature that matters most: it stops spending, it does not just warn you.
- The free plan cannot bill you — it hard-stops when the trial credits are gone.
- Credits reset each month, so a bad week does not follow you into the next.
- Spend alerts and per-run limits mean a single task cannot drain the whole budget.
What a credit actually is
A credit is simply a unit of usage budget, denominated in dollars. When an agent reasons about a task, runs commands, or keeps an app alive, it consumes a small amount of that budget. Credits are the meter; the budget cap is the fuse.
This is different from paying per seat or per project. You are not buying a fixed number of tools. You are buying a pool of compute the agent can spend however the work demands — a little on a quick script, more on a longer build — up to a ceiling you control.
Denominating credits in dollars matters. "You have 40 dollars of agent compute this month" is a sentence a founder or an operator can reason about. "You have 2.7 million tokens" is not. The point of a credit is to make the cost legible to the person approving it.
The budget cap is the feature, not the price
Most pricing pages compete on the headline number. For autonomous agents, the number that actually protects you is the cap.
A budget cap means that when the month's credits are spent, the agent is stopped. Not warned, not throttled after the fact — stopped, in context, while it is working. The agent is told it has hit the limit and cannot continue spending until the budget resets or you raise it. That is what turns "the agent can spend money on its own" from a risk into a bounded, predictable line item.
Every plan carries a cap. What differs is what happens at the ceiling:
- On the free plan, the cap is a hard stop. There is no overage and no bill — when the trial credits are gone, the agent pauses.
- On paid plans, you choose. You can allow billable overage so long-running work is not interrupted, or you can keep a strict cap so spending can never exceed the plan's included credits.
Either way, the decision is yours and it is made in advance, not discovered on an invoice.
Why credits reset monthly
Credits refresh at the start of each billing period. This sounds like a small detail, but it changes how it feels to experiment.
If a single expensive week permanently ate into your budget, you would ration every agent run. A monthly reset means you can have an ambitious week — build three tools, throw two away — and start the next month whole. It rewards experimentation, which is exactly the behavior you want when you are still learning what your agents are good at.
Guardrails below the cap
The cap is the outer fence. Two smaller guardrails keep you from ever getting near it by accident.
The first is spend alerts. As usage approaches the cap, you get an email — not after the fact, but while there is still budget left to act on. You find out an agent is spending faster than expected in time to look, not after the money is gone.
The second is per-run bounds. A single agent run is limited, so no one task can drain the entire monthly budget in one loop. Even a misbehaving agent that gets stuck hits its run limit long before it hits your monthly cap. This is part of what makes it safe to give an agent its own persistent workspace and let it work unattended.
Compute is not the only meter
AI credits cover the agent's reasoning and build work. A couple of other resources are metered separately, and it helps to know which is which:
- Server tiers. Each server is a sized compute box — small, medium, or large — and your plan controls which tiers you can provision. This is the box the agent builds in, priced by capacity.
- Egress. Outbound traffic from your live projects has its own monthly allowance, with predictable per-gigabyte overage above it.
Keeping these separate keeps the mental model clean: credits meter what the agent does, tiers meter how big a box it does it in, and egress meters how much the finished app serves.
How to think about your first month
For a team validating whether agents are worth it, the honest advice is to start small and watch the meter. A modest plan with a hard or low cap lets you answer the only question that matters early — do agents actually produce tools we use? — without betting anything you would miss.
If the answer is yes, you raise the cap. If it is no, you were never exposed. That asymmetry is the whole point of pairing usage credits with a firm budget cap: the downside is fixed and small, and the upside is a library of tools your team actually opens. You can see the current tiers and included credits on the pricing page.
FAQ
What is an AI credit?
A credit is a unit of usage budget, measured in dollars, that the agent spends as it reasons, runs commands, and keeps apps alive. Your plan includes a monthly pool of them.
Can an agent spend more than my budget?
Only if you allow it. Every plan has a budget cap. On the free plan the cap is a hard stop with no possible bill. On paid plans you can enable billable overage or keep a strict ceiling that can never be exceeded.
What happens when the credits run out?
The agent is stopped in context and told it has hit the limit. It resumes when the budget resets at the start of the next month, or sooner if you raise the cap.
Do unused credits roll over?
Credits reset each billing period rather than accumulating, which keeps the monthly cost predictable and rewards experimenting within the period.
Is compute the only thing that is metered?
No. AI credits cover the agent's work. Server tiers (small, medium, large) and outbound egress are metered separately with their own allowances.
Related reading
- Private by default: reviewing agent-built software before you ship it
- Why AI agents need persistent workspaces
- The missing deployment layer for AI agents
- See plans and included credits