# Daytona SDK — Programmable Dev Sandboxes for AI Agents

> Daytona SDK spawns Linux dev environments in 90 ms. Run agent-generated code, browser automation, ML jobs. Snapshot + fork to branch execution.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

1. Sign up at app.daytona.io → copy API key
2. `pip install daytona-sdk` (or `npm install @daytonaio/sdk`)
3. `daytona.create()`, then `sandbox.process.exec("<command>")` — branch with `sandbox.snapshot()`

---

## Intro

The Daytona SDK lets AI agents spin up secure Linux sandboxes in 90 ms. Each sandbox is a full Ubuntu environment with a writable filesystem, optional GPU, and the ability to fork or snapshot to branch agent execution paths. Best for: AI agents that need to test generated code, perform parallel exploratory branches, or maintain a persistent agent workspace. Works with: Python and TypeScript SDKs. Setup time: 5 minutes.

---

### Spin up a sandbox

```python
from daytona import Daytona

daytona = Daytona(api_key=os.environ["DAYTONA_API_KEY"])

sandbox = daytona.create()
print(sandbox.id)  # "sb_abc123"
print(sandbox.cold_start_ms)  # ~90

# Run a command
response = sandbox.process.exec("python -c 'print(2+2)'")
print(response.result)  # "4"

# Or shell session
session = sandbox.process.create_session()
session.execute("cd /workspace && git clone ...")
session.execute("cd repo && pip install -r requirements.txt")
session.execute("pytest")

sandbox.delete()  # or auto-cleanup on exit
```

### Mount files from local

```python
sandbox = daytona.create()

# Upload a directory
sandbox.fs.upload_directory(local_path="./my-project", sandbox_path="/workspace/proj")

# Download generated artifacts
sandbox.fs.download(sandbox_path="/workspace/proj/dist", local_path="./build")
```

### Snapshot + fork — branch agent execution

```python
sandbox = daytona.create()
sandbox.process.exec("git clone https://github.com/example/repo /work")

# Snapshot the current state
snap = sandbox.snapshot()  # "snap_xyz"

# Branch 1: try approach A
sb_a = daytona.create_from_snapshot(snap)
sb_a.process.exec("python try_approach_a.py")

# Branch 2: try approach B in parallel
sb_b = daytona.create_from_snapshot(snap)
sb_b.process.exec("python try_approach_b.py")

# Pick the winner
winner = sb_a if score(sb_a) > score(sb_b) else sb_b
```

### Why use Daytona vs running Docker locally

- 90 ms cold start vs Docker's ~10 s
- Snapshot + fork lets agents do tree search across solutions
- Cloud-resident — no local CPU / RAM tax
- Pay per second of sandbox time

---

### FAQ

**Q: Is Daytona free?**
A: Daytona has a free tier for testing. Paid plans (per-second compute pricing) start when you scale up. Their core platform is also open-source on GitHub for self-hosting.

**Q: How is this different from Modal Sandboxes or E2B?**
A: All three are cloud Linux sandboxes. Differences: Daytona's snapshot/fork primitives are unique (great for agents doing parallel exploration). Modal optimizes for Python ML workloads. E2B optimizes for code-interpreter UX and microVM isolation. Pick by primary use case.

**Q: Can I use Daytona with Claude Code or Cursor?**
A: Yes — agents in Claude Code / Cursor can use the Daytona Python SDK as a tool. They can also connect via MCP — Daytona ships an MCP server. Useful for 'try this code in a sandbox before committing it' workflows.

---

## Source & Thanks

> Built by [Daytona](https://github.com/daytonaio). Apache-2.0 (core).
>
> [daytonaio/daytona](https://github.com/daytonaio/daytona) — ⭐ 16,000+

---

<!-- ZH -->

## 快速使用

1. 在 app.daytona.io 注册，复制 API key
2. `pip install daytona-sdk`（或 `npm install @daytonaio/sdk`）
3. `daytona.create()`，然后 `sandbox.process.exec("<command>")`，用 `sandbox.snapshot()` 分支

---

## 简介

Daytona SDK 让 AI agent 在 90ms 内起一个安全 Linux sandbox。每个 sandbox 是完整 Ubuntu 环境，可写文件系统、可选 GPU、能 fork 或 snapshot 来分叉 agent 执行路径。适合需要测试生成代码、做并行探索分支、或维护持久 agent 工作区的 AI agent。Python 和 TypeScript SDK 都有。装机时间 5 分钟。

---

### 起一个 sandbox

```python
from daytona import Daytona

daytona = Daytona(api_key=os.environ["DAYTONA_API_KEY"])

sandbox = daytona.create()
print(sandbox.id)  # "sb_abc123"
print(sandbox.cold_start_ms)  # ~90

# 跑命令
response = sandbox.process.exec("python -c 'print(2+2)'")
print(response.result)  # "4"

# 或 shell session
session = sandbox.process.create_session()
session.execute("cd /workspace && git clone ...")
session.execute("cd repo && pip install -r requirements.txt")
session.execute("pytest")

sandbox.delete()  # 或退出时自动清理
```

### 从本地挂文件

```python
sandbox = daytona.create()

# 上传目录
sandbox.fs.upload_directory(local_path="./my-project", sandbox_path="/workspace/proj")

# 下载生成的产物
sandbox.fs.download(sandbox_path="/workspace/proj/dist", local_path="./build")
```

### Snapshot + fork —— 分叉 agent 执行

```python
sandbox = daytona.create()
sandbox.process.exec("git clone https://github.com/example/repo /work")

# 快照当前状态
snap = sandbox.snapshot()  # "snap_xyz"

# 分支 1：试方案 A
sb_a = daytona.create_from_snapshot(snap)
sb_a.process.exec("python try_approach_a.py")

# 分支 2：并行试方案 B
sb_b = daytona.create_from_snapshot(snap)
sb_b.process.exec("python try_approach_b.py")

# 选赢的
winner = sb_a if score(sb_a) > score(sb_b) else sb_b
```

### 为啥用 Daytona 不本地跑 Docker

- 90ms 冷启动 vs Docker 的 ~10s
- Snapshot + fork 让 agent 能在解决方案上做树搜索
- 云端 —— 不占本地 CPU / RAM
- 按 sandbox 秒计费

---

### FAQ

**Q: Daytona 免费吗？**
A: Daytona 有免费档用于测试。付费档（按秒计算定价）规模化时启用。核心平台也在 GitHub 开源，可自托管。

**Q: 跟 Modal Sandbox 或 E2B 啥区别？**
A: 三个都是云 Linux sandbox。区别：Daytona 的 snapshot/fork 原语独特（agent 做并行探索时很好）。Modal 优化 Python ML 工作负载。E2B 优化 code-interpreter 体验和 microVM 隔离。按主要用例选。

**Q: 能配 Claude Code 或 Cursor 吗？**
A: 能 —— Claude Code / Cursor 里的 agent 可以把 Daytona Python SDK 当工具用。也能通过 MCP 接入 —— Daytona 出 MCP server。「提交前在 sandbox 里试一下这段代码」这种工作流很合适。

---

## 来源与感谢

> Built by [Daytona](https://github.com/daytonaio). Apache-2.0 (core).
>
> [daytonaio/daytona](https://github.com/daytonaio/daytona) — ⭐ 16,000+


---
Source: https://tokrepo.com/en/workflows/daytona-sdk-programmable-dev-sandboxes-for-ai-agents
Author: Daytona