When sandbox mode is enabled, the Meta Agent can execute code safely in isolated containers. This enables powerful automation while maintaining security.
Required Connector: Sandbox
| Tool | Description |
|---|
execute_python | Run Python code in a sandbox |
execute_shell | Run shell commands in a sandbox |
execute_docker | Run commands in a Docker container |
Python Execution
Run Python code with access to common libraries.
Example Usage:
"Run a Python script to analyze the CSV data"
"Execute Python code to calculate the deployment statistics"
"Parse the JSON response and extract the relevant fields"
Available Libraries
The sandbox includes common Python libraries:
requests - HTTP requestspandas - Data analysisjson - JSON parsingcsv - CSV handlingdatetime - Date/time operations- And many more standard library modules
Shell Execution
Run shell commands in an isolated environment.
Example Usage:
"Run a vulnerability scan on this repository using trivy"
"Execute curl to test the API endpoint"
"Run grep to search for patterns in the logs"
Docker Execution
Run commands in custom Docker containers.
Example Usage:
"Build and test the Docker image"
"Run the test suite in a containerized environment"
"Execute the migration script in a database container"
Sandbox Features
Isolation
- Code runs in ephemeral containers
- Each execution starts with a clean environment
- No persistence between executions
Resource Limits
- Memory: Configurable memory limits
- CPU: CPU time constraints
- Storage: Limited disk space
Network Control
- Configurable network access
- Can restrict outbound connections
- Prevent access to internal networks
Timeout Protection
- Automatic termination of long-running tasks
- Configurable timeout values
- Graceful shutdown handling
File System Isolation
- No access to host file system
- Temporary workspace for each execution
- Automatic cleanup after completion
Security Considerations
While the sandbox provides isolation, always review what code you’re asking the Meta Agent to execute. The sandbox is designed for automation tasks, not for running untrusted code.
Best Practices
- Review generated code before execution for sensitive operations
- Use minimal permissions when configuring the sandbox connector
- Monitor executions through the execution logs
- Set appropriate timeouts to prevent runaway processes
Example Workflows
Data Analysis
User: "Analyze the deployment frequency from this CSV"
Meta Agent: Executes Python to parse CSV, calculate statistics,
and generate a summary report
Security Scanning
User: "Scan the repository for vulnerabilities"
Meta Agent: Runs trivy or similar tools in the sandbox,
parses results, and presents findings
API Testing
User: "Test all endpoints in the OpenAPI spec"
Meta Agent: Generates and executes curl commands or Python
requests to validate API responses
