OpenAI Agents SDK · Blaxel Sandbox

OpenAI agents work, wait, then work again.

Their infrastructure should too. Blaxel keeps the full working state ready through standby, resumes it in 25ms, and drops active-compute cost to $0 while they wait.

Run an OpenAI agent View the Agents SDK guide

Up to $200 in credits · No credit card required

Teams building with Blaxel

OpenAI + Blaxel

OpenAI gives your agent the loop. Blaxel gives it somewhere to continue.

OpenAI Agents SDK handles models, tools, handoffs, guardrails, and traces. Blaxel gives the agent a persistent computer that can pause without losing files, memory, or processes.

Agent layer

OpenAI Agents SDK

Build the intelligence and orchestration of your agent with OpenAI's models, tools, handoffs, guardrails, and traces.

Agent loop and handoffs
Models, tools, and guardrails
Streaming and traces

Native sandbox provider

@openai/agents-extensions

Execution layer

Blaxel Sandbox

Give that agent an isolated computer where commands, files, dependencies, and long-running work can live.

Commands, files, and packages
$0 active compute on standby
25ms resume with state intact
Network and secret controls in preview

Together, your OpenAI agent can

Execute

Run shell commands and generated code in an isolated microVM.

Build

Create files, install dependencies, and move artifacts in and out.

Show its work

Stream output and open a live preview of the application it creates.

Continue

Keep its filesystem, memory, and processes intact on standby, then resume the same work in 25ms.

See an OpenAI agent return to the same workspace.

Up to $200 in credits · No credit card required

Run an OpenAI agent

Native SDK connection

Keep the agent loop. Add the computer.

Blaxel is a built-in sandbox provider for the OpenAI Agents SDK. Choose Blaxel in the agent's run configuration. The SDK handles the remote execution environment.

01
Define the agent in OpenAI Agents SDK
Keep the model, instructions, tools, and run loop in the framework you already use.
02
Connect the Blaxel sandbox provider
Choose the Blaxel provider, image, and workspace root in the run configuration.
03
Let the agent work inside its environment
Commands, files, installed packages, and generated artifacts stay inside the sandbox boundary.

index.ts

TypeScript

$pnpm add @openai/agents-extensions @blaxel/core

1import { run } from "@openai/agents";2import { Manifest, SandboxAgent } from "@openai/agents/sandbox";3import { BlaxelSandboxClient } from4  "@openai/agents-extensions/sandbox/blaxel";5 6const agent = new SandboxAgent({7  name: "Sandboxed agent",8  instructions: "Run commands and work with files.",9  defaultManifest: new Manifest({ root: "/blaxel" }),10});11 12const result = await run(agent, "Run the tests and fix failures.", {13  sandbox: {14    client: new BlaxelSandboxClient(),15    options: { name: "agent-sandbox-01", image: "blaxel/base-image" },16  },17});18 19console.log(result.finalOutput);

OpenAI agent loop · Blaxel execution layer

Open the full Agents SDK guide

Persistence between agent turns

The OpenAI agent returns. Its working state should too.

Blaxel keeps the full sandbox state on standby between actions. Active compute costs $0 while it waits, and the same files, memory, and processes resume in 25ms.

$0 active compute on standby

Keep the work without keeping compute active.

Move the sandbox to standby between agent turns, reviews, or external events. The working state stays ready while active-compute billing stops.

$0 active compute on standby
Filesystem, memory, and processes retained
25ms resume

First-class provider

Connect natively from the Agents SDK.

Use the Blaxel sandbox client directly in the OpenAI Agents SDK run configuration instead of building a custom tool bridge.

TypeScript and Python support
SDK-managed sandbox sessions
Framework-native run configuration

Commands, files, dependencies

Give the agent a complete working environment.

Run commands, manage directories, install packages, execute code, stream output, and move artifacts between the sandbox and your application.

Command and process execution
File transfer and filesystem APIs
Application preview URLs

Concrete access controls

Control what the sandbox can reach.

Use domain filtering and proxy secret injection in public preview. Dedicated egress IPs are available in private preview for stable allowlisting.

Domain filtering in public preview
Proxy secret injection in public preview
Dedicated egress IPs in private preview

Build with the ecosystem

Start with the OpenAI workflow you already have.

Blaxel is built into the OpenAI Agents SDK sandbox-provider ecosystem. Add it where the agent needs a computer, then move from a first command to production-shaped work.

OpenAI ecosystem

OpenAI Agents SDK

Keep the agent, model, instructions, tools, streaming behavior, and run loop in the OpenAI framework.

Native sandbox provider

Blaxel Sandbox

Connect through the Agents SDK extension and add a persistent microVM, filesystem, packages, previews, and network controls.

Connect Blaxel to the OpenAI Agents SDK.

Up to $200 in credits · No credit card required

Run an OpenAI agent

OpenAI Agents SDK sandbox

Questions before you connect an agent.

What is a sandbox for the OpenAI Agents SDK?

A sandbox is the remote computer where an agent can execute commands, work with files, install dependencies, and run generated code. Blaxel is available as a built-in sandbox provider in the OpenAI Agents SDK, so the execution environment is available through the framework's run configuration.

Can an OpenAI agent run shell commands and edit files?

Yes. With a Blaxel Sandbox, the agent can execute commands, inspect processes, create and edit files, transfer artifacts, and run code inside an isolated microVM.

Can the sandbox persist between agent interactions?

Yes. Configure the sandbox lifecycle to pause work on standby instead of deleting the environment. Blaxel retains the filesystem, memory, and running processes, then resumes the same working state in 25ms.

What does $0 active compute on standby mean?

When the sandbox moves to standby, active-compute billing stops. Snapshot storage and any attached volume storage continue to apply while Blaxel preserves the working state for its next 25ms resume.

How do I control secrets and network access?

Domain filtering and proxy secret injection are available in public preview. Dedicated egress IPs are available in private preview when an external system requires a stable allowlisted identity.

Does the integration support TypeScript and Python?

Yes. The OpenAI Agents SDK and Blaxel sandbox integration support both TypeScript and Python. The full guide includes a working example in each language.

OpenAI Agents SDK + Blaxel

Give the OpenAI agent loop a computer it can return to.

Keep files, memory, and processes ready between turns. Active compute costs $0 on standby, and the same sandbox resumes in 25ms.

Run an OpenAI agent