langchain-islo

Islo sandbox integration for Deep Agents.

Islo provides long-running, reconnect-surviving AI sandboxes on real Linux VMs, with pause/resume, snapshots, and a gateway layer for secret isolation and egress policies.

Install

uv add langchain-islo
# or: pip install langchain-islo

Set your API key (keys look like ak_...):

export ISLO_API_KEY="ak_..."

Sandbox image requirement. The inherited filesystem tools (read, write, edit, ls, glob, grep) shell out to python3 and GNU grep inside the sandbox. Use an image that provides them (e.g. python:3.12-slim, or ubuntu/debian with python3 + grep installed). Alpine/BusyBox grep lacks the flags these tools need.

Usage

Wrap an existing Islo sandbox and use it as a Deep Agents backend:

from islo import Islo
from langchain_islo import IsloSandbox

client = Islo()  # reads ISLO_API_KEY
sandbox = client.sandboxes.create_sandbox(image="python:3.12-slim")

backend = IsloSandbox(client=client, sandbox=sandbox)

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

# Filesystem tools are inherited from BaseSandbox:
backend.write("/workspace/app.py", "print('hi')\n")
print(backend.read("/workspace/app.py").file_data["content"])

Lifecycle helper

IsloProvider creates, attaches to, and deletes sandboxes for you:

from langchain_islo import IsloProvider

provider = IsloProvider()  # reads ISLO_API_KEY
backend = provider.get_or_create(image="python:3.12-slim")
try:
    print(backend.execute("uname -a").output)
finally:
    provider.delete(sandbox_id=backend.id)

What you get

IsloSandbox subclasses deepagents.backends.sandbox.BaseSandbox, so it implements the full SandboxBackendProtocol:

Async variants (aexecute, aupload_files, adownload_files, ...) are provided by the base class via thread offloading.

Configuration

License

MIT