axio-tools-docker 0.6.2

Docker sandbox tools for Axio

axio-tools-docker

Docker sandbox tools for axio.

Run agent-generated code and commands inside isolated Docker containers. The agent gets sandbox_exec, sandbox_write, and sandbox_read tools that operate entirely within the sandbox — the host filesystem stays untouched.

Features

• Isolated execution — code runs inside a Docker container, not on the host
• Configurable image — use any Docker image as the sandbox environment
• Three sandboxed tools — execute commands, write files, read files — all inside the container
• Persistent sandbox — container is reused across tool calls within a session for faster execution
• TUI integration — configure image, memory limits, and CPU from the axio-tui settings screen

Requirements

Docker must be installed and running:

docker info   # should succeed

Installation

pip install axio-tools-docker

Usage

from axio.agent import Agent
from axio.context import MemoryContextStore
from axio_transport_openai import OpenAITransport
from axio_tools_docker.plugin import DockerPlugin

async def main() -> None:
    plugin = DockerPlugin()
    await plugin.init()   # uses default config (python:3.12-slim)

    agent = Agent(
        system=(
            "You are a coding assistant. Use sandbox_exec to run code safely. "
            "Never attempt to access the host filesystem directly."
        ),
        tools=plugin.all_tools,
        transport=OpenAITransport(api_key="sk-...", model="gpt-4o"),
    )

    ctx = MemoryContextStore()
    result = await agent.run(
        "Write a Python script that computes the first 20 Fibonacci numbers and run it.",
        ctx,
    )
    print(result)

Sandbox tools

Tool	Description
sandbox_exec	Run a shell command inside the container; returns stdout + stderr
sandbox_write	Write a file into the container's filesystem
sandbox_read	Read a file from the container's filesystem

Container lifecycle

The sandbox container is created lazily on the first tool call (sandbox_exec, sandbox_write, or sandbox_read) — plugin.init() itself does not start Docker. The same container is reused for all subsequent tool calls within the session.

When await plugin.close() is called the container is stopped and removed (docker rm -f). You should always call close() when the agent session ends to avoid leaving containers behind:

async def run_with_cleanup(agent, ctx, plugin):
    try:
        result = await agent.run("...", ctx)
    finally:
        await plugin.close()

If Docker is not installed or the docker CLI is not on PATH, the container creation step will raise a RuntimeError on the first tool call. You can detect this before starting a session with SandboxManager.docker_available() (returns True if the docker executable is found on PATH).

Configuration

Field	Type	Default	Description
image	str	"python:latest"	Docker image to use for the sandbox container
memory	str	"256m"	Memory limit passed to docker run --memory (e.g., "512m", "1g")
cpus	str	"1.0"	CPU limit passed to docker run --cpus
network	bool	False	Whether to allow network access. When False, --network none is set on the container.
workdir	str	"/workspace"	Working directory inside the container

from axio_tools_docker.config import SandboxConfig

config = SandboxConfig(
    image="python:3.12-slim",
    memory="512m",
    cpus="1.0",
    workdir="/workspace",
)

Plugin registration

[project.entry-points."axio.tools.settings"]
docker = "axio_tools_docker.plugin:DockerPlugin"

Part of the axio ecosystem

axio · axio-tools-local · axio-tools-mcp · axio-tui

License

MIT