langchain-sandbox-opensandbox

OpenSandbox sandbox integration for Deep Agents.

Quick Install

uv add langchain-sandbox-opensandbox

The distribution is published as langchain-sandbox-opensandbox, but the import package is langchain_opensandbox.

from opensandbox import SandboxSync

from langchain_opensandbox import OpenSandboxBackend

sandbox = SandboxSync.create("python:3.12")
backend = OpenSandboxBackend(sandbox=sandbox, timeout=300)

result = backend.execute("echo hello")
print(result.output)

Using it with Deep Agents

Pass the backend to a Deep Agent so the agent's file and shell tools run inside the OpenSandbox environment instead of on the host:

from deepagents import create_deep_agent
from opensandbox import SandboxSync

from langchain_opensandbox import OpenSandboxBackend

# 1. Spin up a sandbox and wrap it.
sandbox = SandboxSync.create("python:3.12")
backend = OpenSandboxBackend(sandbox=sandbox, timeout=300)

# 2. Hand the backend to the agent. Every ls / read_file / write_file /
#    glob / grep / execute the agent performs is now sandboxed.
agent = create_deep_agent(
    tools=[],
    system_prompt="You are a coding assistant. Use the sandbox to run code.",
    backend=backend,
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Create hello.py and run it."}]}
)
print(result["messages"][-1].content)

# 3. Tear the sandbox down when you're done.
sandbox.kill()

The higher-level file helpers the agent calls (ls, read_file, write_file, glob, grep) are provided by BaseSandbox and built on top of the three primitives this adapter implements (execute, upload_files, download_files).

🤔 What is this?

OpenSandboxBackend adapts the OpenSandbox Python SDK to the BaseSandbox interface used by Deep Agents, so you can run agent-generated commands and file operations inside an OpenSandbox environment.

It implements the three sandbox primitives — execute, upload_files, and download_files — on top of the OpenSandbox SDK. The higher-level file helpers (ls, read_file, write_file, glob, grep) are provided by BaseSandbox and built on top of execute.

Configuration

OpenSandboxBackend wraps an existing opensandbox.SandboxSync instance, so it inherits whatever connection you configured on the SDK. To point at a specific server, build a ConnectionConfigSync and pass it to SandboxSync.create:

from opensandbox import SandboxSync
from opensandbox.config.connection_sync import ConnectionConfigSync

from langchain_opensandbox import OpenSandboxBackend

connection = ConnectionConfigSync(
    domain="127.0.0.1:8080",  # host (and optional port) of the server
    protocol="http",          # "http" or "https"
    api_key="my-api-key",     # if the server requires authentication
)
sandbox = SandboxSync.create("python:3.12", connection_config=connection)
backend = OpenSandboxBackend(sandbox=sandbox)

Running a local OpenSandbox server

The integration tests (and the snippets above) need a reachable OpenSandbox server. You can run one locally with Docker:

# Requires Docker to be running.
uvx opensandbox-server init-config ./os-sandbox.toml --example docker
uvx opensandbox-server --config ./os-sandbox.toml

This starts the management API on 127.0.0.1:8080. Set an api_key in the generated config if you want authentication; otherwise the server starts in insecure mode. Then point the SDK at it with domain="127.0.0.1:8080" and protocol="http" as shown above.

Development

uv sync
make lint            # ruff check + format --diff
make test            # unit tests (no network)
make integration_tests   # conformance suite (requires a running OpenSandbox server)
make build           # build wheel + sdist

Conformance

The integration tests run the standard SandboxIntegrationTests conformance suite from langchain-tests — the same suite every Deep Agents sandbox backend is validated against. It exercises execute, file upload/download, and all the inherited helpers (ls, read_file, write_file, glob, grep) with real I/O, including large payloads, escaped content, error handling, and both the sync and async paths.

This adapter passes the full suite (86/86) against a live OpenSandbox server — no mocks, no skips:

OPENSANDBOX_DOMAIN=127.0.0.1:8080 \
OPENSANDBOX_PROTOCOL=http \
OPENSANDBOX_API_KEY=<your-key> \
make integration_tests

The suite is skipped automatically unless OPENSANDBOX_DOMAIN is set, so make test stays offline and fast.

Releases & Versioning

This package follows semantic versioning.

Contributing

Contributions are welcome. Please open an issue or pull request on GitHub.

License

Apache License 2.0. See LICENSE.