Sandboxed: A Secure Code Execution Environment

sandboxed.md

Why Sandboxed?

So we have built a self hostable language-agnostic framework to execute code in isolated environments for your AI agents.

It uses your existing kubernetes cluster to spin up isolated sandboxes for executing code securely. Each sandbox is ephemeral and destroyed after use, ensuring no residual data or state persists.

https://github.com/system32-ai/sandboxed

how-it-works.md

How does it work?

Agents can invoke sandboxed environments to run code securely and independently. The sandboxes are created in the kubernetes that's integrated with the system.

The lifecycle of these sandboxes is managed automatically - they are created when an agent needs to execute code, and destroyed when the task is complete. This allows for efficient resource usage and scalability.

Primitive to control the sandboxes are provided via APIs, allowing agents to request sandbox creation, execute code within them, and retrieve results.

example.md

Using API

curl -L https://github.com/system32-ai/sandboxed/releases/download/v1.0.8/sandboxed-v1.0.0-darwin-amd64.tar.gz -o
sandboxed.tar.gz
tar -xzf sandboxed.tar.gz
mv sandboxed-v1.0.8-darwin-arm64 sandboxed
sudo mv sandboxed /usr/local/bin/
sudo chmod +x /usr/local/bin/sandboxed
rm sandboxed.tar.gz
sandboxed version

Start the sandboxed server:

sandboxed server

Create a sandbox, execute code and destroy it:

Create a sandbox

curl -X POST http://localhost:8080/api/v1/sandbox/create \
-H "Content-Type: application/json" \
-d '{
    "language": "python",
    "name": "data-processing",
}'

Execute code in the sandbox

curl -X POST http://localhost:8080/api/v1/execute \
-H "Content-Type: application/json" \
-d '{
    "sandbox_id": "data-processing-1698765432",
    "code": "import pandas as pd\nimport numpy as np\n\n# Create sample data\ndata = np.random.randn(100, 3)\ndf =
    pd.DataFrame(data, columns=[\"A\", \"B\", \"C\"])\nprint(f\"Dataset shape: {df.shape}\")\nprint(f\"Mean
    values:\\n{df.mean()}\")",
}'

Destroy the sandbox

curl -X POST http://localhost:8080/api/v1/sandbox/destroy \
-H "Content-Type: application/json" \
-d '{
    "sandbox_id": "data-processing-1698765432",
    "namespace": "development",
    "force": false
}'

Tool's list in MCP

Tool	Purpose	Parameters	Returns
`create_sandbox`	Create new sandbox	name, language, namespace?, labels?	success, message
`run_code`	Execute code in sandbox	sandbox_name, code	success, output, exit_code
`destroy_sandbox`	Clean up sandbox	sandbox_name	success, message
`list_sandboxes`	List active sandboxes	none	sandboxes[], count

Using Native SDK

Here's a simple example of how an AI agent can use the sandboxed environment to execute Python code securely using the SDK:


package main

import (
    "log"
    "github.com/system32-ai/sandboxed/pkg/sdk"
)

func main() {

    sandbox, err := sdk.CreateSandbox("debug-generated-code", sdk.Python)
    if err != nil {
    log.Fatalf("failed to create sandbox: %v", err)
    }

    defer sandbox.Destroy()

    code := `python -c 'print("Hello, World!")'`

    output, err := sandbox.Run(code)
    if err != nil {
    log.Fatal(err)
    }

    log.Printf("Output: %s", output.Result)

    code = `python --version`
    output, err = sandbox.Run(code)
    if err != nil {
    log.Fatal(err)
    }

    log.Printf("Output: %s", output.Result)
}

═══════════════════════════════════════════════════════════════════════════════════

Back to Blogs