Now in early access

Secure sandboxes
for AI agents

Run Codex, Claude Code, OpenCode, and custom AI agents inside isolated, pause-resumable cloud sandboxes. Built for server-side agents that need to execute code, edit files, run tools, expose ports, and keep long-running workspaces safe.

Powered by CubeSandbox Open-source runtime VM-backed isolation

AI agents are no longer just chatbots. They clone repos, run tests, install packages, call CLIs, start dev servers, browse files, and operate inside real environments.

Runra Sandbox gives every agent a secure, disposable, observable execution environment — without requiring you to operate VMs, Kubernetes, KVM, networking, snapshots, or sandbox infrastructure yourself.

Powered by CubeSandbox. Integrated with the open-source Runra runtime.

What is Runra Sandbox?

Hosted sandbox infrastructure for AI agents. Isolated cloud workspaces where agents can safely run code, use tools, modify files, start processes, expose ports, pause, resume, and terminate.

Your AppRunra RuntimeRunra SandboxCubeSandbox-backed isolated execution environments

Use it directly through the API, or plug it into the open-source Runra runtime as your default sandbox backend.

Built for agent workloads

Traditional serverless platforms are designed for short-lived functions. AI agents need something different.

Long-running tasks
📁 Mutable workspaces
💾 File system state
Code execution
📦 Package installation
🔄 Background processes
🌐 Dev servers
🔒 Network controls
⏯️ Pause and resume
📊 Full execution traces

Runra Sandbox is designed for these agent-native workloads from the beginning.

Run agents safely

Give each user, task, or workspace its own isolated sandbox. Run untrusted code, generated scripts, package installs, repo operations, and shell commands without exposing your main application infrastructure.

VM-backed sandbox execution

Hardware-level isolation with full kernel separation between sandboxes.

Workspace-level file system isolation

Each sandbox gets its own file system. No cross-workspace access.

Process isolation

Processes in one sandbox cannot interact with processes in another.

Resource limits

CPU, memory, and storage limits applied per sandbox.

Network policy controls

Fine-grained network access rules for outbound connections.

Timeout and lifecycle controls

Automatic termination based on timeout and lifecycle policies.

Disposable environments

Sandboxes are designed to be ephemeral and disposable.

Pause/resume support

Preserve workspace state without keeping environments active.

Pause and resume agent workspaces

Agents often work in sessions, not requests. Runra Sandbox lets you preserve workspace state across agent runs without keeping every environment fully active forever.

  • Long-running coding agents
  • User-specific development workspaces
  • Multi-step research agents
  • Background automation agents
  • Stateful tool-using agents
  • Cost-controlled warm environments
▶️
Active sandbox
Running · Consuming runtime
⏸️
Paused sandbox
State preserved · No active runtime
▶️
Resumed sandbox
State restored · Ready to execute

Works with your agent stack

Runra is designed to work with different agents, models, tools, and observability systems.

Agent adapters

  • Claude Code
  • Codex
  • OpenCode
  • Custom agents
  • CLI-based agents
  • Tool-calling agents

LLM providers

  • OpenAI
  • Anthropic
  • Gemini
  • OpenRouter
  • Local models
  • Custom providers

Observability exporters

  • Axiom
  • OpenTelemetry
  • JSONL
  • Postgres
  • Custom exporters

Sandbox providers

  • Runra Sandbox
  • CubeSandbox self-hosted
  • Docker
  • E2B
  • Custom providers

Runra Sandbox is the hosted default, but the runtime stays open and pluggable.

Sandbox created
Sandbox started
Command executed
Tool invoked
File changed
Port exposed
Sandbox paused
Sandbox resumed
Sandbox terminated
Runtime errors
Resource usage
Duration and cost signals

Observability-friendly by default

Runra Sandbox emits structured lifecycle and execution events that can be sent to your own observability stack.

Bring your own Axiom, OpenTelemetry, Postgres, ClickHouse, or log pipeline. No black-box agent execution.

Why Runra Sandbox?

🚫

Do not run agents on your app servers

Agent-generated code, shell commands, package installs, and background processes should not run next to your production application. Run them in isolated sandboxes instead.

🏗️

Do not build sandbox infrastructure from scratch

Operating secure, scalable sandbox infrastructure requires VM lifecycle management, networking, scheduling, snapshotting, resource limits, API gateways, and fleet management. Runra Sandbox handles this layer for you.

🔌

Do not lock into one agent framework

Runra is open and adapter-based. Use the agents, LLM providers, observability tools, and sandbox backends that fit your stack.

🚀

Start hosted, self-host later

Use Runra Sandbox when you want managed infrastructure. Use CubeSandbox directly when you want to self-host the underlying sandbox engine.

Use cases

How teams are using Runra Sandbox for agent workloads.

Coding agents

Run Claude Code, Codex, OpenCode, or custom coding agents in isolated repo workspaces. Agents can clone repos, edit files, run tests, install dependencies, and expose preview servers.

Data and analytics agents

Let agents run scripts, inspect files, generate reports, execute notebooks, and operate in controlled environments.

Workflow automation agents

Give background agents safe execution environments for recurring or long-running tasks.

Multi-tenant SaaS agents

Create one sandbox per user, workspace, organization, or task to isolate customer data and execution state.

Agent development platforms

Build your own agent platform without building the sandbox layer from scratch.

API example

Create a sandbox, run commands, expose a port, and pause — all through the SDK.

example.ts
import { Runra } from "@runra/sdk";

const runra = new Runra({
  sandbox: {
    provider: "runra-sandbox",
    apiKey: process.env.RUNRA_API_KEY,
  },
  observability: {
    provider: "axiom",
    token: process.env.AXIOM_TOKEN,
    dataset: "runra-events",
  },
});

const sandbox = await runra.sandboxes.create({
  image: "node:22",
  resources: {
    cpu: 2,
    memoryMb: 4096,
  },
});

await sandbox.exec("git clone https://github.com/acme/app.git");
await sandbox.exec("cd app && npm install && npm test");

const port = await sandbox.exposePort(3000);

await sandbox.pause();

Open-source runtime.
Hosted sandbox.

Runra is open source. You can use it to standardize how your agents run, observe, control, and connect to different infrastructure providers.

Runra Sandbox is the hosted sandbox backend for teams that do not want to operate sandbox infrastructure themselves.

Open-source

  • Runtime
  • SDK
  • Agent adapters
  • LLM provider adapters
  • Observability exporters
  • Sandbox provider interface

Hosted

  • Runra Sandbox API
  • Managed sandbox fleet
  • Scheduling
  • Pause/resume
  • Resource limits
  • Billing
  • Operational reliability

Powered by

CubeSandbox

Runra Sandbox is backed by CubeSandbox, a VM-based sandbox engine designed for fast, isolated, stateful execution environments.

CubeSandbox provides the lower-level sandbox engine. Runra Sandbox provides the hosted API, lifecycle management, scheduling, and developer experience for agent workloads.

Build agents that can safely run real work

Use Runra Sandbox to give every AI agent a secure, isolated, pause-resumable execution environment.