# Modal Sandboxes — Secure Cloud Code Execution for AI Agents

> Modal Sandboxes spin up secure Linux environments for agent-generated code in seconds. Custom images, GPUs, persistent volumes from any Modal Function.

## Install

Copy the content below into your project:

## Quick Use

1. `pip install modal && modal token new`
2. Define an image with `modal.Image.debian_slim().pip_install(...)`
3. Spawn `modal.Sandbox.create(image=...)` and call `sb.exec(...)`

---

## Intro

Modal Sandboxes are isolated Linux environments you spin up programmatically — perfect for running agent-generated code, fine-tuning, or untrusted user code. Custom Docker-style images, optional GPUs, persistent volumes, sub-second cold start. Best for: AI agents that need to execute generated Python / shell / Node code safely. Works with: Modal SDK (Python). Setup time: 5 minutes.

---

### Spawn a sandbox

```python
import modal

app = modal.App("agent-sandbox")
image = modal.Image.debian_slim(python_version="3.12").pip_install("pandas", "numpy", "matplotlib")

@app.function(image=image)
def run_agent_code(code: str):
    sb = modal.Sandbox.create(
        image=image,
        cpu=2.0,
        memory=4096,
        timeout=300,
        app=app,
    )

    process = sb.exec("python", "-c", code)
    stdout = process.stdout.read()
    stderr = process.stderr.read()
    sb.terminate()

    return {"stdout": stdout, "stderr": stderr, "exit_code": process.returncode}
```

### Sandbox with persistent storage

```python
volume = modal.Volume.from_name("agent-workspace", create_if_missing=True)

sb = modal.Sandbox.create(
    image=image,
    volumes={"/workspace": volume},
    workdir="/workspace",
)

# Files written to /workspace persist across sandboxes
sb.exec("python", "fetch_data.py")
volume.commit()

# Later sandbox reads the same files
sb2 = modal.Sandbox.create(volumes={"/workspace": volume})
sb2.exec("python", "process_data.py")
```

### GPU sandbox

```python
sb = modal.Sandbox.create(
    image=image,
    gpu="A10G",  # or "H100", "T4", etc
    timeout=3600,
)
sb.exec("python", "train.py")
```

### Why use Modal Sandboxes vs Docker locally

- Spin up in <1s vs Docker's tens of seconds
- Pre-built images cached at the platform level
- Pay per second of execution
- Same SDK works for Functions, Sandboxes, Volumes, GPUs
- Built-in monitoring, logs, dashboards

---

### FAQ

**Q: Is Modal free?**
A: Modal has a free tier ($30/mo platform credit). Beyond that you pay per second of compute (CPU + memory + GPU). No platform fee — just resources used.

**Q: How is this different from E2B?**
A: E2B and Modal Sandboxes overlap heavily. E2B optimizes for sandbox-as-a-product (the fastest cold start, isolated networking by default). Modal optimizes for the bigger Modal platform (Functions, Volumes, queues, GPUs). Both work for agent code execution; pick by ecosystem fit.

**Q: Can I install custom packages in the sandbox?**
A: Yes — define an image with `modal.Image.debian_slim().pip_install(...)` or `apt_install(...)` or `dockerfile_commands(...)`. The image is built once and cached for fast cold starts.

---

## Source & Thanks

> Built by [Modal](https://github.com/modal-labs). Commercial product with free tier.
>
> [modal.com/docs](https://modal.com/docs/guide/sandbox) — Sandbox documentation

---

<!-- ZH -->

## 快速使用

1. `pip install modal && modal token new`
2. 用 `modal.Image.debian_slim().pip_install(...)` 定义镜像
3. 起 `modal.Sandbox.create(image=...)` 然后调 `sb.exec(...)`

---

## 简介

Modal Sandbox 是程序化启动的隔离 Linux 环境 —— 适合跑 agent 生成的代码、微调、不可信用户代码。自定义 Docker 风格镜像、可选 GPU、持久 volume、亚秒级冷启动。适合需要安全执行生成的 Python / shell / Node 代码的 AI agent。需要 Modal SDK（Python）。装机时间 5 分钟。

---

### 起一个 sandbox

```python
import modal

app = modal.App("agent-sandbox")
image = modal.Image.debian_slim(python_version="3.12").pip_install("pandas", "numpy", "matplotlib")

@app.function(image=image)
def run_agent_code(code: str):
    sb = modal.Sandbox.create(
        image=image,
        cpu=2.0,
        memory=4096,
        timeout=300,
        app=app,
    )

    process = sb.exec("python", "-c", code)
    stdout = process.stdout.read()
    stderr = process.stderr.read()
    sb.terminate()

    return {"stdout": stdout, "stderr": stderr, "exit_code": process.returncode}
```

### 带持久存储的 sandbox

```python
volume = modal.Volume.from_name("agent-workspace", create_if_missing=True)

sb = modal.Sandbox.create(
    image=image,
    volumes={"/workspace": volume},
    workdir="/workspace",
)

# 写到 /workspace 的文件跨 sandbox 保留
sb.exec("python", "fetch_data.py")
volume.commit()

# 后面的 sandbox 读同样的文件
sb2 = modal.Sandbox.create(volumes={"/workspace": volume})
sb2.exec("python", "process_data.py")
```

### GPU sandbox

```python
sb = modal.Sandbox.create(
    image=image,
    gpu="A10G",  # 或 "H100" / "T4" 等
    timeout=3600,
)
sb.exec("python", "train.py")
```

### 为啥用 Modal Sandbox 而不是本地 Docker

- 启动 <1s，Docker 要几十秒
- 预构建镜像在平台层缓存
- 按执行秒计费
- 同一 SDK 涵盖 Function / Sandbox / Volume / GPU
- 自带监控、日志、仪表盘

---

### FAQ

**Q: Modal 免费吗？**
A: Modal 有免费档（每月 $30 平台 credit）。之后按计算秒（CPU + 内存 + GPU）收费。没有平台费 —— 只算资源用量。

**Q: 跟 E2B 啥区别？**
A: E2B 和 Modal Sandbox 高度重叠。E2B 优化「沙箱即产品」（最快冷启动、默认隔离网络）。Modal 优化整个 Modal 平台（Function / Volume / 队列 / GPU）。都能跑 agent 代码；按生态适配选。

**Q: 能在 sandbox 里装自定义包吗？**
A: 能 —— 用 `modal.Image.debian_slim().pip_install(...)` 或 `apt_install(...)` 或 `dockerfile_commands(...)` 定义镜像。镜像构建一次，缓存以加速冷启动。

---

## 来源与感谢

> Built by [Modal](https://github.com/modal-labs). Commercial product with free tier.
>
> [modal.com/docs](https://modal.com/docs/guide/sandbox) — Sandbox documentation


---
Source: https://tokrepo.com/en/workflows/modal-sandboxes-secure-cloud-code-execution-for-ai-agents
Author: Modal