Your Agents Can Provision Infrastructure. Your Budget Can't Run Away.
Handing an AI agent the keys to provision its own infrastructure is the biggest productivity jump in agentic engineering, and the biggest fear. Here is the part of Open Source Cloud that makes it a safe jump instead of a leap of faith: every agent, human or AI, spends from the same pool of tokens, and when that pool runs low, provisioning stops and waits for a person rather than billing you for the difference.
I wrote earlier this week about the four stages of agentic engineering, and about how the last one, letting agents provision infrastructure and not just write code, is the stage most engineers stall in front of. Not because the technology isn't there. Because handing an agent a database, a storage bucket, or a transcoding pipeline feels like handing it a blank check. That reaction is correct, if the infrastructure underneath is a normal cloud bill. An agent with real AWS credentials and a bad prompt can spin up resources all weekend before a human notices. That fear is the actual trust limit, more than the accountability question I wrote about last time. Nobody wants to explain a five-figure invoice caused by an agent that was just trying things. OSC doesn't ask you to trust the agent not to overspend. It makes a runaway bill structurally impossible.
Tokens are a ceiling, not a meter
Every service instance on OSC, whether a human provisions it through the console or an agent provisions it through MCP, draws from the same token balance. Tokens are consumed by running infrastructure: a database that stays up, a transcoding job that runs, a storage bucket that holds data. They are not consumed by prompts, by AI reasoning, or by how many times an agent iterates before it gets something right. Pay for the infrastructure, not for the AI's thought process. That distinction matters here specifically because it means the ceiling is on infrastructure, the exact thing an autonomous agent is now allowed to provision, not on how much the agent is allowed to think or attempt.
What the ceiling actually looks like
Every plan comes with a token allowance, and that allowance is the hard stop. Basic is free: 100 one-time tokens, no card required. Personal, at 15 EUR a month, gives 40 tokens a day plus a 300-token bonus on upgrade, and adds database backups and scheduled agent tasks. Professional, at 69 EUR a month, gives 300 tokens a day plus the same 300-token bonus, and adds an IMAP and SMTP mail server. Business, at 199 EUR a month, gives 1,000 tokens a day. Enterprise is custom, for teams that have outgrown the fixed tiers. All prices exclude VAT. The tiers differ mainly by token volume rather than by feature gating: custom domains, team collaboration, unlimited hosted applications and an SLA are included on every plan, Basic included.
What happens when the tokens run out
Your daily allowance is added to your balance each day, and unused tokens carry forward rather than expiring, so the balance behaves like a running budget rather than a use-it-or-lose-it quota. When that balance runs low, the platform stops rather than quietly billing you for the difference. In the console, creating an instance you cannot afford is blocked and you are asked to top up first. Over MCP, an agent cannot provision silently: the first call returns a cost estimate alongside your current balance, and the agent has to come back and explicitly confirm before anything is created. Topping up is a one-off token pack, Starter at 9.90 EUR for 100 tokens or Builder at 25 EUR for 300 tokens, both valid for 12 months, or an upgrade to a larger daily allowance. That is a deliberate design choice, not a missing feature. The moment infrastructure spend needs a decision, it stops and waits for a human to make one.
Back to stage 4
In the stage-4 argument, I said engineers hang on to code review out of identity as much as risk, and that non-engineers reach full autonomy faster because they never built that identity in the first place. Cost is where that same hesitation shows up again, just dressed as a technical concern instead of an emotional one. I can't let an agent touch the database is sometimes a considered position. More often, in my experience, it's a trust limit wearing a technology costume. A fixed token ceiling doesn't ask you to resolve that tension by trusting the agent more. It resolves it by making the blank check impossible to write. You can hand over provisioning the way you'd hand over a corporate card with a real limit on it, not an open line of credit.
See the numbers
The full breakdown of tiers and token packs, including what a token actually buys you per service, is on the pricing page.
Try it
Connect your agent to OSC and see what it can provision inside your plan's fixed ceiling.
Frequently Asked Questions
Can an agent spend beyond my plan's daily allowance?
A balance can technically go negative. What cannot happen is a bill for it. The console blocks creating an instance you do not have the balance for, and over MCP an agent has to explicitly confirm the cost before anything is provisioned. Beyond that there is no overage and no surprise invoice: you are never billed beyond your subscription and any token packs you choose to buy. On the free plan, an account that runs far enough past empty has its instances suspended rather than invoiced.
Do tokens work differently for AI agents than for a person using the console?
No. Tokens are consumed by the infrastructure itself, not by who or what requested it. An agent and a human draw from the same balance under the same rules.
Can I set a lower ceiling than my plan allows, for a specific project or agent team?
Not as a separate limit today. The daily or one-time token allowance for the account is the ceiling every agent operating under it shares. If you want a harder boundary for a specific team, a lower-tier account with its own smaller allowance is the current way to do that.
Do unused daily tokens roll over?
Yes. The daily refill is added to your balance and unused tokens carry forward, so the balance accumulates rather than resetting each day. Purchased token packs do expire, 12 months from the date of purchase.
Does this apply to My Apps and My Agent Tasks too, or just service instances?
All of it. Anything that runs on OSC infrastructure, deployed by a person, a scheduled agent task, or an autonomous agent team, draws from the same token balance.
Related Posts
Your Marketing Team Doesn't Need a Developer. It Needs an Agent.
One person with no engineering background closed the gap between "the video is edited" and "the video is published, captioned, and reviewed" in about a day. Here is what he built, and why it only worked because the AI agent had real infrastructure to call, not just code to write.
Your Agent Doesn't Just Deploy the Backend. It Can Run It.
Most agent-infrastructure stories stop at "the agent deployed it." A real run this week went further: an agent uploaded, packaged, and delivered a review copy, then reported an operational mistake nobody asked it to check. Here is why that only works when the agent is calling real, unmodified open-source APIs.
Give your AI agent a backend it can deploy, and you can walk away with
Your AI coding agent can build the app. OSC gives it a backend to deploy through one MCP connection, built from unmodified open-source services you can export and run yourself, on any cloud, any time.