Framework integrations

Adapters for the agent frameworks that have a sandbox interface.

QuickstartIf you have not run one yet BudgetsCaps that hold across adapters ReferenceThe underlying client

LangChain deepagents

pip install langchain-opsen

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

backend = OpsenSandbox(task_id="nightly-report", budget_usd=2.00)

agent = create_deep_agent(
    model=ChatAnthropic(model="claude-sonnet-4-6"),
    system_prompt="You are a Python coding assistant.",
    backend=backend,
)

backend.cost()   # machine + model + tools, one task id

Subclasses BaseSandbox, so the filesystem tools and async variants come from deepagents itself. Passes 79 of 86 checks in their sandbox standard suite.

OpenAI Agents SDK

pip install openai-agents-opsen

from agents import Runner, RunConfig
from agents.sandbox import SandboxAgent
from openai_agents_opsen import OpsenSandboxClient

result = await Runner.run(
    agent,
    "read the repo and summarise it",
    run_config=RunConfig(sandbox={"client": OpsenSandboxClient(
        task_id="code-review", budget_usd=2.00)}),
)

VibeKit

npm install @vibe-kit/opsen

import { VibeKit } from "@vibe-kit/sdk";
import { createOpsenProvider } from "@vibe-kit/opsen";

const vibeKit = new VibeKit()
  .withAgent({ type: "grok", provider: "xai",
               apiKey: process.env.XAI_API_KEY! })
  .withSandbox(createOpsenProvider({
    apiKey: process.env.OPSEN_API_KEY!,
    taskId: "code-review",
    budgetUsd: 2.0,
  }));

Pydantic AI

pip install pydantic-ai-opsen

from pydantic_ai import Agent
from pydantic_ai_opsen import OpsenSandbox

agent = Agent(
    "anthropic:claude-sonnet-4-6",
    capabilities=[OpsenSandbox(task_id="nightly-report",
                               budget_usd=2.00)],
)

A capability, so it needs no upstream registration — you add it where you build the agent. It carries four tools: run a command, read, write, list.

Temporal

SandboxClientProvider("opsen", OpsenSandboxClient(
    task_id="nightly-report", budget_usd=2.00))

No Temporal-specific package: their integration takes any BaseSandboxClient, and the Agents SDK adapter is one. See Temporal.

agentsh

pip install agentsh-opsen

from agentsh_secure_sandbox import secure_sandbox, SecureConfig
from agentsh_opsen import OpsenAdapter

sandbox = await secure_sandbox(
    OpsenAdapter(task_id="review", budget_usd=2.00),
    SecureConfig(install_strategy="running"))

agentsh wraps a sandbox in runtime security and exposes it to Pydantic AI, LangChain, the OpenAI Agents SDK and deepagents. It needs its own daemon inside the sandbox, so this wants an image with agentsh preinstalled — the adapter is ready, that image is not.

TypeScript, directly

npm install @opsendev/opsen

import { Opsen } from "@opsendev/opsen";

const opsen = new Opsen({ apiKey: process.env.OPSEN_API_KEY });
const s = await opsen.createSession({ taskId: "x", budgetUsd: 1 });
const r = await s.exec("echo hello");
await s.kill();

From your editor

opsen speaks MCP, so Claude Code, Cursor and VS Code can start runs without leaving them:

claude mcp add --transport http --scope user opsen \
  https://opsen.dev/mcp \
  --header "Authorization: Bearer YOUR_OPSEN_KEY"

PreviousCommand line NextMCP server