Google ADK + OpenSandbox Example

Integrate Google Agent Development Kit (ADK) with OpenSandbox. The agent uses ADK's native OpenSandboxEnvironment and EnvironmentToolset to execute commands and edit files inside a sandbox. Requires Python 3.11+ and Google ADK 2.13.0 or newer.

Experimental API

ADK currently marks the OpenSandbox environment and environment toolset as experimental.

Start OpenSandbox server [local]

Pre-pull the code-interpreter image (includes Python):

docker pull sandbox-registry.cn-zhangjiakou.cr.aliyuncs.com/opensandbox/code-interpreter:v1.1.0

# use docker hub
# docker pull opensandbox/code-interpreter:v1.1.0

Start the local OpenSandbox server, logs will be visible in the terminal:

uv pip install opensandbox-server
opensandbox-server init-config ~/.sandbox.toml --example docker
opensandbox-server

Run the example

The agent uses DeepSeek through ADK's LiteLlm adapter. Install the connector and its MCP runtime dependency:

uv pip install "google-adk[opensandbox,extensions]>=2.13.0" "mcp>=2.2,<3"

export OPENAI_API_KEY="your-deepseek-api-key"
export SANDBOX_DOMAIN="localhost:8080"
# Set SANDBOX_API_KEY too if your server requires authentication

# Run from the repository root
uv run python examples/google-adk/main.py

The default model is openai/deepseek-flash at https://api.deepseek.com/v1. To use another OpenAI-compatible endpoint, set GOOGLE_ADK_MODEL with the openai/ prefix and set OPENAI_BASE_URL. The endpoint must support tool calling.

The script writes /workspace/math.py, executes it to produce 5754, then writes and reads back ADK + OpenSandbox from /workspace/notes.txt. All prompts share one session and workspace. Tool events and model responses are printed to the terminal.

Native ADK integration

from google.adk.integrations.opensandbox import OpenSandboxEnvironment
from google.adk.tools.environment import EnvironmentToolset
from opensandbox.config import ConnectionConfig

toolset = EnvironmentToolset(
    environment=OpenSandboxEnvironment(
        image="sandbox-registry.cn-zhangjiakou.cr.aliyuncs.com/opensandbox/code-interpreter:v1.1.0",
        timeout=300,
        connection_config=ConnectionConfig(domain="localhost:8080"),
    )
)
# Pass tools=[toolset] to the ADK Agent.
# In your async application's finally block: await toolset.close()

ADK provides Execute, WriteFile, ReadFile, and EditFile. The toolset initializes the environment lazily. Relative paths resolve under /workspace; commands also run there. The environment creates a sandbox with a 300-second lifetime and renews it before command execution. The example closes the toolset in finally, destroying the sandbox it created even when a request fails.

Earlier versions of this example wrapped the OpenSandbox SDK in three custom functions. The example now uses ADK's native environment integration while retaining the same entry point and sandbox environment variables. It explicitly passes ConnectionConfig, so the existing SANDBOX_* names continue to work.

Environment Variables

Variable	Default	Description
SANDBOX_DOMAIN	localhost:8080	Sandbox service address
SANDBOX_API_KEY	(optional for local)	API key if your server requires authentication
SANDBOX_IMAGE	sandbox-registry.cn-zhangjiakou.cr.aliyuncs.com/opensandbox/code-interpreter:v1.1.0	Sandbox image to use
GOOGLE_ADK_MODEL	openai/deepseek-flash	Model name with the openai/ prefix
OPENAI_API_KEY	(required)	API key for the selected endpoint
OPENAI_BASE_URL	https://api.deepseek.com/v1	OpenAI-compatible endpoint base URL

References

• Google ADK - Agent Development Kit

• Source code on GitHub

• ADK 2.13.0 release

• ADK native OpenSandbox sample