E2B — Cloud Sandboxes for AI Code Execution

What it is

E2B provides secure, isolated cloud sandboxes where AI agents and LLM applications can execute code safely. Each sandbox is a lightweight microVM that spins up in approximately 150ms, runs arbitrary code (Python, JavaScript, shell), and tears down when done. The sandbox includes a filesystem, package manager, and network access.

E2B targets developers building AI agents, coding assistants, or any LLM application that needs to execute user-provided or AI-generated code without risking the host system.

How it saves time or tokens

E2B eliminates the need to build and maintain your own code execution infrastructure. Without it, you would set up Docker containers, manage security isolation, handle process timeouts, and build a REST API around it. E2B provides all of this as a service with a simple SDK.

The 150ms startup time means code execution feels instant in user-facing applications. There is no cold start penalty that would make an AI assistant feel sluggish.

How to use

1. Install the E2B SDK:

npm install @e2b/code-interpreter

1. Create a sandbox and run code:

import { Sandbox } from '@e2b/code-interpreter';

const sandbox = await Sandbox.create();

const result = await sandbox.runCode('print(sum(range(100)))');
console.log(result.text); // 4950

await sandbox.close();

1. Install packages and use the filesystem:

await sandbox.runCode('!pip install pandas');
await sandbox.runCode(`
import pandas as pd
df = pd.DataFrame({'a': [1,2,3], 'b': [4,5,6]})
print(df.describe())
`);

Example

Integrating E2B with an AI agent for code execution:

import { Sandbox } from '@e2b/code-interpreter';
import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();
const sandbox = await Sandbox.create();

// AI generates code, E2B executes it
const message = await client.messages.create({
  model: 'claude-sonnet-4-20250514',
  messages: [{ role: 'user', content: 'Write Python to plot sin(x)' }],
  max_tokens: 1024,
});

const code = extractCode(message.content);
const result = await sandbox.runCode(code);
console.log(result.text);

await sandbox.close();

Related on TokRepo

• AI tools for agents -- Platforms and tools for building AI agents
• AI tools for coding -- AI code execution and development tools

Common pitfalls

• Not setting execution timeouts. AI-generated code can contain infinite loops. Always set a timeout on sandbox.runCode() to prevent runaway processes.
• Forgetting to close sandboxes. Each sandbox consumes resources. Call sandbox.close() in a finally block or use try-with-resources patterns to avoid orphaned sandboxes.
• Assuming network access is always safe. Sandboxes have internet access by default. If your use case involves untrusted code, consider restricting network access to prevent data exfiltration.