opsen documentation

Run agents on any runtime, with machine, model and tool spend under one task id.

Quickstart An agent on a real machine in about two minutes, with the cost of that run printed at the end. Sessions The unit everything hangs off: one machine, one token, one line on the bill. Frameworks LangChain deepagents, the OpenAI Agents SDK, VibeKit, MCP — and the raw clients. Budgets and caps A ceiling the gateway enforces mid-run, rather than an alert about money already spent.

What opsen does

Your agent runs on a sandbox — E2B, Modal, or your own machines — and its model calls are proxied through the same session that owns that sandbox. That is the whole design, and everything else follows from it.

Because opsen sits in both paths, one number covers the machine, the model and the tools your agent called:

{'total_usd': 0.41, 'compute_usd': 0.02, 'tokens_usd': 0.39,
 'calls': 34, 'task_id': 'nightly-report'}

And because the proxy is in the request path rather than reading a log afterwards, a budget is a refusal rather than an alert: an agent that would cross its cap is stopped mid-run, not discovered on an invoice.

Explore

Concepts Sessions One machine, one token, one line on the bill Isolation What separates agent code from your secrets Runtimes E2B, Modal, or your own machines The gateway Why the model calls go through opsen

Running agents Python client Sessions, files, exec and cost Command line For pipelines and for running something now MCP server Start runs from Claude Code or Cursor Files Reading and writing the sandbox

Cost control Budgets and caps Refused mid-run, not reported after Choosing a model Name one, or say what matters Which model is better On your runs, with the honesty to say when Balance and billing Prepaid, top-up, and what stops

Security Agent credentials Your GitHub, used by an agent that never holds it Egress control Where a sandbox may talk to API keys Authenticating, and what to do when one leaks Tasks and batches One id, many runs, one ceiling

Your agent needs no opsen code

The sandbox starts with the provider environment variables already set, so any SDK inside it routes through opsen with no configuration:

import anthropic

client = anthropic.Anthropic()   # no api_key, no base_url
msg = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=400,
    messages=[{"role": "user", "content": "hello"}],
)

What is injected is a session token, not a provider key: scoped to one session, capped by that session's budget, dead when the session ends. A long-lived provider key inside a machine running generated code is the thing this design exists to avoid. See Environment for the full list.

NextQuickstart