langchain-opsen

Run LangChain deepagents on opsen — with machine, model and tool spend on one bill and a budget the runtime enforces mid-run.

pip install langchain-opsen

from deepagents import create_deep_agent
from langchain_anthropic import ChatAnthropic
from langchain_opsen import OpsenSandbox

agent = create_deep_agent(
    model=ChatAnthropic(model="claude-sonnet-4-6"),
    system_prompt="You are a Python coding assistant.",
    backend=OpsenSandbox(task_id="nightly-report", budget_usd=2.00),
)

Set OPSEN_API_KEY, or pass api_key=. Get one here:

https://opsen.dev/keys

What this adds

Every sandbox backend bills you for the machine and leaves the model bill with a different vendor, on a different invoice, with no shared identifier. Two bills, reconciled by hand, and no answer to "what did that agent cost".

opsen sits in both paths, so:

agent_cost = backend.cost()
# {'total_usd': 0.41, 'compute_usd': 0.02, 'tokens_usd': 0.39,
#  'calls': 34, 'input_tokens': 84120, 'output_tokens': 9310,
#  'task_id': 'nightly-report'}

budget_usd is a ceiling the gateway enforces during the run. An agent that would cross it is refused mid-flight rather than noticed afterwards on an invoice.

Behaviour worth knowing

The session starts lazily. Constructing OpsenSandbox starts no machine — the first command does. An agent built and never run costs nothing.

File operations and the shell share one channel. Both go through the same exec path, so the file tools and execute() always agree about the working directory.

Ports depend on the runtime. host(port) works on E2B, works on Modal when the port was declared at sandbox creation, and raises on the local runtime with a message naming the limitation. A provider that hands back a URL which does not answer turns "preview my server" into a debugging session.

Configuration

Licence

Apache-2.0