langchain-withruntime

Runtime sandbox integration for Deep Agents.

Runtime runs each sandbox in a Firecracker microVM with its own kernel. RuntimeSandbox is a Deep Agents sandbox backend: the agent's execute, ls, read_file, write_file, edit_file, glob and grep tools all run in one sandbox.

Quick install

pip install langchain-withruntime

This installs the Runtime SDK, withruntime, and Deep Agents. Set RUNTIME_API_KEY to a key from https://withruntime.com/account/keys, or run npx -y withruntime login once on this machine. A new account starts on a free trial with no card.

Use it

from deepagents import create_deep_agent
from withruntime import Sandbox

from langchain_withruntime import RuntimeSandbox

with Sandbox.create() as sbx:
    agent = create_deep_agent(
        model="anthropic:claude-sonnet-4-6",
        backend=RuntimeSandbox(sbx),
    )
    result = agent.invoke(
        {"messages": [{"role": "user", "content": "Write fib.py, run it, and show the first 10 numbers."}]}
    )
    print(result["messages"][-1].content)

Leaving the with block stops the sandbox. The backend works on its own too:

from withruntime import Sandbox

from langchain_withruntime import RuntimeSandbox

with Sandbox.create() as sbx:
    backend = RuntimeSandbox(sbx, timeout_seconds=300)
    print(backend.execute("python3 --version").output)
    backend.write("/workspace/notes.txt", "hello\n")
    print(backend.read("/workspace/notes.txt").file_data["content"])

• timeout_seconds (default 1800) is the limit for a command that names none; a command that runs past it is stopped and its output says so.
• max_output_chars (default 100,000) keeps the end of a longer output and marks it truncated.
• Sandbox.create() takes the sandbox's settings: image, region, vcpu, memory_mib, timeout_seconds (its lease), network and more. See the Python SDK guide.

Deep Agents Code

The package registers a runtime sandbox provider for Deep Agents Code:

dcode install langchain-withruntime --package
dcode --sandbox runtime

Each session gets a new sandbox, stopped when the session ends. --sandbox-id <id> attaches to a sandbox that is already running and leaves it running. RuntimeProvider is the same lifecycle in code:

from langchain_withruntime import RuntimeProvider

provider = RuntimeProvider()
backend = provider.get_or_create(timeout=900)  # a new sandbox with a 15 minute lease
try:
    print(backend.execute("uname -r").output)
finally:
    provider.delete(sandbox_id=backend.id)

Where it comes from

RuntimeSandbox is maintained in the Runtime SDK as withruntime.deepagents (pip install "withruntime[deepagents]" gives the same class). This package installs it under the langchain-<provider> name the other Deep Agents sandbox packages use and adds the Deep Agents Code provider.

Tests

pip install -e . pytest pytest-asyncio langchain-tests
pytest tests/unit_tests                         # offline
RUNTIME_API_KEY=... pytest tests/integration_tests  # real sandboxes
RUNTIME_API_KEY=... python scripts/e2e.py           # a Deep Agent with a scripted model

The offline tests run LangChain's standard SandboxIntegrationTests against a stand-in sandbox that executes on the local machine (Linux).

Documentation

• LangChain, LangGraph and Deep Agents on Runtime
• Runtime documentation

License

Apache-2.0