dvm-run

dvm.run's Python SDK.

pip install dvm-run
dvm-run login

dvm-run login opens dvm.run in your browser, where you approve an API key for this machine; it saves the key in ~/.config/dvm/credentials.json, where the SDK finds it. Over SSH, open the link it prints wherever you like. Where no one is there to approve one, such as CI, make a key on dvm.run's API keys page and set DVM_API_KEY.

A sandbox is a VM made from an image: your agent sends it commands and files until it stops it. Everything it ran is its execution, which replays, the same, once it ends.

from dvm import Sandbox, Execution

with Sandbox.create("python:3.12", seed=7) as sbx:
    sbx.files.write("/app/main.py", "print(1 + 1)\n")
    r = sbx.commands.run("python main.py", cwd="/app", timeout=60)
    print(r.stdout, r.exit_code)

    server = sbx.commands.run(["python", "-m", "http.server"], background=True)
    server.kill()
    server.wait()

verdict = Execution.get(sbx.id).replay().wait()
assert verdict.same

It takes dvm.run's address as api_url=, else DVM_API_URL, and its key as api_key=, else DVM_API_KEY, else the one dvm-run login saved for that address. AsyncSandbox and AsyncExecution are the same API for asyncio.

• A string runs in the sandbox's shell (/bin/bash -lc, or /bin/sh -c in an image without bash); a list runs as it is.
• A command's exit code is never an exception. A sandbox that ended raises SandboxEndedError (SandboxFailedError, with dvm's reason, if it failed) on the next call.
• timeout= kills the command once it passes; its result says so, with code 137 and signal 9.