Secure sandboxes for AI agents

Run code without trusting the code.

Runra Sandbox gives every agent an isolated, pause-resumable cloud workspace for code execution, files, package installs, ports, and long-running tasks.

quickstart.ts
import { Sandbox } from "runra";

const sandbox = await Sandbox.create({ timeout: 300 });

try {
  const result = await sandbox.runCode(
    'print("hello from Runra")'
  );
  console.log(result.text);
} finally {
  await sandbox.kill();
}

Agent-native infrastructure

Built for workloads that do more than chat.

Agents clone repos, run tests, install packages, start servers, inspect files, and need stateful workspaces.

01

Long-running tasks

Run agents through multi-step jobs, tests, package installs, notebooks, and background processes.

02

Mutable workspaces

Each sandbox has a real filesystem that can clone repos, write files, and preserve state.

03

Pause and resume

Stop paying for active runtime while keeping workspace state ready for the next agent turn.

04

Port previews

Expose dev servers and generated apps through controlled public URLs.

05

VM isolation

Run generated code in tenant-isolated environments backed by CubeSandbox.

06

E2B-compatible SDK

Use Runra-native API keys while keeping familiar sandbox execution APIs.

Safety model

One sandbox per user, task, or workspace.

Keep untrusted shell commands, generated code, and package installs away from your application runtime with VM-backed process and filesystem isolation.

Your App
  -> Runra API
  -> Runra Sandbox
  -> CubeSandbox-backed isolated execution environment

Integrations

Designed for modern agent stacks.

Claude CodeCodexOpenCodeOpenAIAnthropicGeminiAxiomOpenTelemetryCubeSandboxE2B-compatible tools