opendsec

Async Python client for opendsec: open-source Sandboxes for RL rollouts.

Each trial gets a World: a Sandbox for the agent's workspace and one for every service the task needs, on one private network, started and torn down as one.

pip install opendsec

import asyncio

import opendsec


async def main():
    # Reads OPENDSEC_URL (default http://localhost:7700) and OPENDSEC_TOKEN.
    async with opendsec.Client() as client:
        world = await client.create_world(
            [
                opendsec.Sandbox(name="env", image="docker.io/library/python:3.12-slim",
                                 command=["python", "-m", "http.server", "8080"],
                                 aliases=["svc.internal"]),
                opendsec.Sandbox(name="main", image="docker.io/library/python:3.12-slim",
                                 command=["sleep", "infinity"]),
            ],
            ttl="5m",
        )
        result = await world.exec("main", "getent hosts svc.internal")
        print(result.exit_code, result.stdout)
        await world.stop()


asyncio.run(main())

• Client.create_world(sandboxes, ttl) returns once every Sandbox is ready, or raises WorldFailed if the platform couldn't deliver the World.
• Client.build_image(dir) builds an image on the platform from a Dockerfile and its build context, and returns once it is ready (a lookup if those files were built before). Use its ref as a Sandbox's image. No Docker needed; only you can run it.
• World.exec, upload and download act in one Sandbox, by name.
• get_world and list_worlds look Worlds up.

Run Harbor rollouts on the hosted test bed: https://opendsec.vibrantlabs.com

Licensed under Apache-2.0.