Run agents anywhere.
One API for any sandbox or your own machines, with model, compute and tool spend on one bill.
A git repo or a pasted script. We read the files, work out how to install it, and run it.
Or you do. Either way you see which one, what it costs per million tokens, and how much context it has — before it runs.
The machine, the model and the tools your agent called, added up per task. Nobody else can show you that, because nobody else is in all three paths.
What it replaces
Today an agent that does real work needs a sandbox provider, a model provider, a few tool APIs, and a way to know what any of it cost. That is four accounts, four bills, and a spreadsheet to reconcile them.
opsen is one account. Top up a balance and everything draws from it.
If you have not built one yet
You do not need to choose a model, a machine, or a provider. Paste a repository URL and press run. We pick sensible defaults, show you what we picked, and let you change any of it later.
If you have
- Sessions with hard budgets. A runaway agent is refused mid-flight, not discovered on an invoice.
- Cost attributed per task id, split into machine, model and tools.
- Bring your own provider keys if you want your own billing and rate limits.
- Name any model, or let us choose by cost, balance or capability.
- Python and TypeScript clients. Your agent needs no opsen code — the sandbox is started with model access already wired in.
The bit worth saying plainly
We hold the provider relationship so you do not have to. That means we are paying the model bill first and recovering it from your balance, which is why budgets and caps are real limits rather than reports.
If you would rather pay providers directly, bring your own key. Nothing about the runtime changes.