deepagents-sbx (Python)

Docker Sandboxes (sbx) microVM sandbox backend for Deep Agents.

This package implements BaseSandbox on top of a Docker Sandboxes microVM — its own kernel plus a private Docker daemon — so agents run arbitrary code behind a hard isolation boundary.

Install

pip install deepagents-sbx            # backend only
pip install "deepagents-sbx[code]"    # + Deep Agents Code provider

Requires Python 3.12+ and the free sbx CLI installed and logged in (sbx login). Local sandboxes always need the CLI; for Docker Cloud Sandboxes you can skip it with the opt-in API transport.

Use

from deepagents import create_deep_agent
from deepagents_sbx import SbxSandbox

with SbxSandbox(memory="4g") as backend:          # creates + auto-removes
    agent = create_deep_agent(model="anthropic:...", backend=backend)
    agent.invoke({"messages": [{"role": "user", "content": "Run the test suite"}]})

Bind-mount a host project (mounted at the same absolute path inside the VM):

with SbxSandbox(workspace="/path/to/project") as backend:
    ...

Cloud via the API (optional)

Cloud normally uses the same CLI with --cloud. To drive Docker Cloud Sandboxes over Docker's experimental Sandboxes API with no sbx process, swap in ApiSbxTransport (standard library only, no new dependency):

from deepagents_sbx import ApiSbxTransport, SbxSandbox

transport = ApiSbxTransport(docker_id="you", personal_access_token="dckr_pat_…")
with SbxSandbox(cloud=True, transport=transport, ttl="10m") as backend:
    backend.execute("echo hello")

Auth is independent of sbx login (OAuth or a PAT with the sandbox:use permission). The API is experimental and cloud-only.

Deep Agents Code

Installing the code extra registers the sbx provider:

dcode --sandbox sbx

See the repository README for the full architecture, API mapping, and testing strategy.

License

MIT