LangGraph Custom Code Interpreter Example

This example demonstrates using a custom code interpreter with PUBLIC network mode, allowing code to access external APIs and internet resources.

Features

Custom Code Interpreter: Creates a CodeInterpreterCustom CloudFormation resource
PUBLIC Network Mode: Code can access external APIs (unlike default SANDBOX mode)
Custom Tool Wrapper: Uses CodeInterpreter directly with custom identifier
Network Validation: Test script verifies PUBLIC mode by fetching external data

Network Mode Comparison

Mode	Network Access	Use Case
SANDBOX (default)	None	Maximum security, local computation
PUBLIC (this example)	Internet	Fetch APIs, download data
VPC	Private VPC	Access internal resources

Project Structure

langgraph-code-interpreter-custom/
├── serverless.yml     # Custom interpreter + agent configuration
├── agent.py           # LangGraph agent with custom interpreter
├── Dockerfile         # Container definition
├── pyproject.toml     # Python dependencies
├── test-invoke.py     # Validation script (tests PUBLIC mode)
└── README.md          # This file

Prerequisites

AWS account with Bedrock model access
Docker installed
Serverless Framework v4+
Python 3.14+

Deployment

Deploy the agent:

serverless deploy

Note the runtime endpoint URL from the output (the ARN is embedded in the URL path):

ai:
  codeInterpreters: N/A
  agents:
    codeAgent: https://bedrock-agentcore.us-east-1.amazonaws.com/runtimes/arn:aws:bedrock-agentcore:us-east-1:123456789012:runtime/.../invocations

Testing

Run the validation script to verify PUBLIC network mode works:

export RUNTIME_ARN="arn:aws:bedrock-agentcore:us-east-1:123456789012:runtime/..."
python test-invoke.py

The test asks the agent to fetch data from GitHub's public API. This would fail in SANDBOX mode but succeeds in PUBLIC mode.

How It Works

Custom Interpreter Configuration

ai:
  codeInterpreters:
    publicInterpreter:
      description: Code interpreter with public internet access
      network:
        mode: PUBLIC # Enables external network access

Agent Code

from bedrock_agentcore.tools.code_interpreter_client import CodeInterpreter

# Get custom interpreter ID from environment
CUSTOM_INTERPRETER_ID = os.environ.get("CUSTOM_INTERPRETER_ID")

# Create client with custom identifier
code_interpreter = CodeInterpreter(region=AWS_REGION)
code_interpreter.start(identifier=CUSTOM_INTERPRETER_ID)  # Uses PUBLIC mode

# Execute code (can access external APIs)
response = code_interpreter.invoke(
    method="executeCode",
    params={"code": code, "language": "python"}
)

code_interpreter.stop()

IAM Permissions

The agent's role needs permission to use the custom interpreter:

role:
  statements:
    - Effect: Allow
      Action:
        - bedrock-agentcore:InvokeCodeInterpreter
      Resource: !GetAtt PublicInterpreterCodeInterpreter.CodeInterpreterArn

Example Prompts

# Fetch external API data (requires PUBLIC mode)
"Fetch the current Bitcoin price from a public API"

# Download and process data
"Download the iris dataset from scikit-learn and show summary statistics"

# Web scraping
"Fetch the HTML from example.com and count the number of paragraphs"

Cleanup

Remove the deployed resources:

serverless remove

langgraph-code-interpreter - Basic example with default interpreter (SANDBOX)

Features

Custom Code Interpreter: Creates a CodeInterpreterCustom CloudFormation resource

PUBLIC Network Mode: Code can access external APIs (unlike default SANDBOX mode)

Custom Tool Wrapper: Uses CodeInterpreter directly with custom identifier

Network Validation: Test script verifies PUBLIC mode by fetching external data

Mode

Network Access

Use Case

SANDBOX (default)

None

Maximum security, local computation

PUBLIC (this example)

Internet

Fetch APIs, download data

VPC

Private VPC

Access internal resources

Project Structure

langgraph-code-interpreter-custom/
├── serverless.yml     # Custom interpreter + agent configuration
├── agent.py           # LangGraph agent with custom interpreter
├── Dockerfile         # Container definition
├── pyproject.toml     # Python dependencies
├── test-invoke.py     # Validation script (tests PUBLIC mode)
└── README.md          # This file

Deployment

Deploy the agent:

serverless deploy

Note the runtime endpoint URL from the output (the ARN is embedded in the URL path):

ai:
  codeInterpreters: N/A
  agents:
    codeAgent: https://bedrock-agentcore.us-east-1.amazonaws.com/runtimes/arn:aws:bedrock-agentcore:us-east-1:123456789012:runtime/.../invocations

How It Works

ai:
  codeInterpreters:
    publicInterpreter:
      description: Code interpreter with public internet access
      network:
        mode: PUBLIC # Enables external network access

from bedrock_agentcore.tools.code_interpreter_client import CodeInterpreter

# Get custom interpreter ID from environment
CUSTOM_INTERPRETER_ID = os.environ.get("CUSTOM_INTERPRETER_ID")

# Create client with custom identifier
code_interpreter = CodeInterpreter(region=AWS_REGION)
code_interpreter.start(identifier=CUSTOM_INTERPRETER_ID)  # Uses PUBLIC mode

# Execute code (can access external APIs)
response = code_interpreter.invoke(
    method="executeCode",
    params={"code": code, "language": "python"}
)

code_interpreter.stop()

The agent's role needs permission to use the custom interpreter:

role:
  statements:
    - Effect: Allow
      Action:
        - bedrock-agentcore:InvokeCodeInterpreter
      Resource: !GetAtt PublicInterpreterCodeInterpreter.CodeInterpreterArn

# Fetch external API data (requires PUBLIC mode) "Fetch the current Bitcoin price from a public API" # Download and process data "Download the iris dataset from scikit-learn and show summary statistics" # Web scraping "Fetch the HTML from example.com and count the number of paragraphs"