Code Interpreter

The run_code tool enables Letta agents to execute code in a secure sandboxed environment. Useful for data analysis, calculations, API calls, and programmatic computation.

Quick Start

1
from letta import Letta
2
3
client = Letta(token="LETTA_API_KEY")
4
5
agent = client.agents.create(
6
    model="openai/gpt-4o",
7
    tools=["run_code"],
8
    memory_blocks=[{
9
        "label": "persona",
10
        "value": "I can run Python code for data analysis and API calls."
11
    }]
12
)

Tool Parameters

Return Format

1
{
2
  "results": ["Last expression value"],
3
  "logs": {
4
    "stdout": ["Print statements"],
5
    "stderr": ["Error output"]
6
  },
7
  "error": "Error details if execution failed"
8
}

Output types:

• results[]: Last expression value (Jupyter-style)
• logs.stdout: Print statements and standard output
• logs.stderr: Error messages
• error: Present if execution failed

Supported Languages

Python

Full Python ecosystem with common packages pre-installed:

• Data: numpy, pandas, scipy, scikit-learn
• Web: requests, aiohttp, beautifulsoup4
• Utilities: matplotlib, PyYAML, Pillow

Check available packages:

1
import pkg_resources
2
print([d.project_name for d in pkg_resources.working_set])

JavaScript & TypeScript

No npm packages available - only built-in Node modules.

1
// Works
2
const fs = require('fs');
3
const http = require('http');
4
5
// Fails
6
const axios = require('axios');

R

Base R only - no tidyverse packages.

1
# Works
2
mean(c(1, 2, 3))
3
4
# Fails
5
library(ggplot2)

Java

JShell-style execution - statement-level only.

1
// Works
2
System.out.println("Hello");
3
int x = 42;
4
5
// Fails
6
public class Main {
7
    public static void main(String[] args) { }
8
}

Network Access

The sandbox has full network access for HTTP requests, API calls, and DNS resolution.

1
import requests
2
3
response = requests.get('https://api.github.com/repos/letta-ai/letta')
4
data = response.json()
5
print(f"Stars: {data['stargazers_count']}")

No State Persistence

Variables, files, and state do not carry over between executions. Each run_code call is completely isolated.

1
# First execution
2
x = 42
3
4
# Second execution (separate run_code call)
5
print(x)  # Error: NameError: name 'x' is not defined

Implications:

• Must re-import libraries each time
• Files written to disk are lost
• Cannot build up state across executions

Self-Hosted Setup

For self-hosted servers, configure an E2B API key. E2B provides the sandbox infrastructure.

$
docker run \
>
  -e E2B_API_KEY="your_e2b_api_key" \
>
  letta/letta:latest

Common Patterns

Data Analysis

1
agent = client.agents.create(
2
    model="openai/gpt-4o",
3
    tools=["run_code"],
4
    memory_blocks=[{
5
        "label": "persona",
6
        "value": "I use Python with pandas and numpy for data analysis."
7
    }]
8
)

API Integration

1
agent = client.agents.create(
2
    model="openai/gpt-4o",
3
    tools=["run_code", "web_search"],
4
    memory_blocks=[{
5
        "label": "persona",
6
        "value": "I fetch data from APIs using run_code and search docs with web_search."
7
    }]
8
)

Statistical Analysis

1
agent = client.agents.create(
2
    model="openai/gpt-4o",
3
    tools=["run_code"],
4
    memory_blocks=[{
5
        "label": "persona",
6
        "value": "I perform statistical analysis using scipy and numpy."
7
    }]
8
)

When to Use

Related Documentation

• Utilities Overview
• Web Search
• Fetch Webpage
• Custom Tools
• Tool Variables