Mastering AI Code Execution in Secure Sandboxes with E2B

Running AI-generated code in production or evaluation environments often comes with major security and infrastructure challenges. Whether you’re building intelligent agents or automated data workflows, securely executing dynamic code is no easy task. Enter E2B, an open-source infrastructure that provides lightning-fast, isolated sandboxes for safely executing AI-generated code in the cloud. In this article, we’ll take a deep dive into how E2B works, explore its powerful Python SDK, and walk you through a step-by-step implementation for real-world applications like AI-driven data analysis.

Table of Content

1. What is E2B?
2. How E2B Works
3. Key Features
4. Step-by-Step Guide

Let’s start by understanding what E2B is.

What is E2B?

E2B (Execute to Build) is an open-source tool providing isolated and secure cloud sandboxes for running AI-generated code. These sandboxes function as small virtual machines, specifically designed for the safe evaluation and execution of code produced by language models. This setup enables users to confidently work with code from models like LLaMA 3 and Code LLaMA.

The versatility of E2B extends to various applications, including evaluating codegen models, building intelligent coding agents capable of reasoning and acting, visualizing datasets analyzed by AI, and running automatically generated applications such as Fragments. Interaction with E2B is facilitated through convenient SDKs available in both Python and JavaScript, offering flexibility for diverse development workflows.

How E2B Works

Beneath the surface, each E2B Sandbox operates as a nimble virtual machine that initializes in approximately 150 milliseconds, delivering swift and scalable runtime environments. These sandboxes are hosted securely in the cloud and maintain complete isolation, ensuring the safety of executed code.

Furthermore, sandboxes can be allocated on a per-user or per-session basis, offering granular control over resource management. They provide essential functionalities such as file operations, the ability to install Python packages via pip, and network access, effectively serving as a personalized coding VM designed to handle individual LLM requests or user sessions.

Key Features

Here are some standout features of E2B:

• Pre-installed Data Libraries: Pandas, Matplotlib, and more, ready out of the box.
• Full Isolation: Prevents unsafe code from affecting your systems.
• Fast Spin-Up: Sandboxes start in under 200ms.
• File Upload/Download Support: Manage datasets or results easily.
• LLM Compatibility: Works seamlessly with models from Together AI or OpenAI.

E2B Key Features

These features make it ideal for advanced AI workflows where reliability, security, and flexibility are critical.

Step-by-Step Guide

This tutorial walks you through how to:

• Use LLaMA-3 70B on Together AI to generate Python code.
• Execute that code securely inside a sandboxed E2B Code Interpreter.
• Upload a dataset and visualize relationships.

Step 1. Install Required Packages

%pip install together==1.3.1 e2b_code_interpreter==1.0.0 python-dotenv==1.0.0

Step 2. Import Libraries and Load Environment Variables

import os
from dotenv import load_dotenv
import json
import re
from together import Together
from e2b_code_interpreter import Sandbox

load_dotenv()

Step 3. Setup Your API Keys

E2B_API_KEY = "e2b_XXXX"  # Replace with your real E2B key
TOGETHER_API_KEY = "together_XXXX"  # Replace with your Together API key

Step 4. Choose a Model from Together AI

MODEL_NAME = "meta-llama/Llama-3.3-70B-Instruct-Turbo-Free"

You can explore other models at: https://api.together.ai/models.  Together AI provides $1 in free credits, which you can use to test various models. Some models, like LLaMA-3.3-70B-Instruct-Turbo-Free, are available for free usage as well.

Step 5. Define the System Prompt

SYSTEM_PROMPT = """You're a Python data scientist. You are given tasks to complete and you run Python code to solve them.

Information about the csv dataset:

- It's in the `/Users/aniruddha/Documents/data.csv` file
- The CSV file is using , as the delimiter
- It has the following columns (examples included):
    - country: "Argentina", "Australia"
    - Region: "SouthAmerica", "Oceania"
    - Surface area (km2): for example, 2780400…. ""

This prompt tells the LLM how to behave and what the dataset looks like. Here you have to add all the column names as it is in dataset.

Step 6. Build Helper Function to Run Code

def code_interpret(e2b_code_interpreter, code):
    print("Running code interpreter...")
    exec = e2b_code_interpreter.run_code(code,
        on_stderr=lambda stderr: print("[Code Interpreter]", stderr),
        on_stdout=lambda stdout: print("[Code Interpreter]", stdout))

    if exec.error:
        print("[Code Interpreter ERROR]", exec.error)
    else:
        return exec.results

Step 7. Extract Python Code Block from LLM Response using Regex

pattern = re.compile(r"```python\n(.*?)\n```", re.DOTALL)

def match_code_blocks(llm_response):
    match = pattern.search(llm_response)
    if match:
        code = match.group(1)
        print(code)
        return code
    return ""

Step 8. Send Chat Prompt to LLaMA + Execute Code if Found

def chat_with_llm(e2b_code_interpreter, user_message):

    print(f"\n{'='*50}\nUser message: {user_message}\n{'='*50}")
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": user_message},
    ]

    response = client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
    )

    response_message = response.choices[0].message
    python_code = match_code_blocks(response_message.content)

    if python_code != "":
        code_interpreter_results = code_interpret(e2b_code_interpreter, python_code)
        return code_interpreter_results

    else:
        print(f"Failed to match any Python code in model's response {response_message}")

        return []

Step 9. Upload Dataset to the Code Interpreter

def upload_dataset(code_interpreter):
    print("Uploading dataset to Code Interpreter sandbox...")
    dataset_path = "/content/data.csv"

    if not os.path.exists(dataset_path):
        raise FileNotFoundError("Dataset file not found")

    try:
        with open(dataset_path, "rb") as f:
            code_interpreter.files.write(dataset_path, f)
        print("Uploaded at", dataset_path)

        return dataset_path

    except Exception as error:
        print("Error during file upload:", error)
        raise error

Step 10. Run Everything in a Secure Sandbox

client = Together(api_key=TOGETHER_API_KEY)

with Sandbox(api_key=E2B_API_KEY) as code_interpreter:
    upload_dataset(code_interpreter)
    code_results = chat_with_llm(
        code_interpreter,

        "Create a bar chart showing Unemployment (% of labour force) for 10 countries with the highest GDP per capita (current US$).",

    )

    if code_results:
        first_result = code_results[0]

    else:
        raise Exception("No code interpreter results")

# View the result
first_result

Output

The model will:

• Parse the dataset.
• Filter missing or invalid values.
• Generate and execute a linear regression chart showing the relationship between GDP per capita and life expectancy.

Final Thoughts

E2B acts as a game-changer for AI-native development. It will be helpful whether you’re working on autonomous agents, code generation tasks, or complex evaluation pipelines, its sandbox infrastructure removes friction and risk. Paired with LLMs from Together AI or others, E2B allows you to safely prototype and test advanced AI workflows in minutes. You can check their official docs and GitHub examples to dive deeper into custom applications.

References

• E2B Official GitHub Repo
• E2B Documentation
• Together AI Models