VercelSandbox integration

Integrate with the VercelSandbox sandbox backend using LangChain Python.

Vercel Sandbox provides ephemeral, isolated Linux environments for running untrusted code. See the Vercel Sandbox docs for signup, authentication, and platform details.

Installation

pip

pip install langchain-vercel-sandbox

uv add langchain-vercel-sandbox

Authentication

The Vercel SDK reads credentials from the environment. Set the following variables, or use OIDC when running on Vercel:

export VERCEL_TOKEN="your-token"
export VERCEL_PROJECT_ID="your-project-id"
export VERCEL_TEAM_ID="your-team-id"

Create a sandbox backend

In Python, you create the sandbox using the provider SDK, then wrap it with the deepagents backend.

from vercel.sandbox import Sandbox

from langchain_vercel_sandbox import VercelSandbox

sandbox = Sandbox.create()
backend = VercelSandbox(sandbox=sandbox)

try:
    result = backend.execute("echo hello")
    print(result.output)
finally:
    sandbox.stop()

Use with Deep Agents

from vercel.sandbox import Sandbox
from langchain_anthropic import ChatAnthropic

from deepagents import create_deep_agent
from langchain_vercel_sandbox import VercelSandbox

sandbox = Sandbox.create()
backend = VercelSandbox(sandbox=sandbox)

agent = create_deep_agent(
    model=ChatAnthropic(model="claude-sonnet-4-20250514"),
    system_prompt="You are a coding assistant with sandbox access.",
    backend=backend,
)

try:
    result = agent.invoke(
        {"messages": [{"role": "user", "content": "Create a small Python project and run tests"}]}
    )
finally:
    sandbox.stop()

Cleanup

Always stop the sandbox when you are done to avoid ongoing resource usage.

See also: Sandboxes.

---

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.

Edit this page on GitHub or file an issue.