Skip to content
Isorun Docs

LangChain

Terminal window
pip install isorun[langchain]

The integration ships two tools:

  • IsorunCodeInterpreterTool — runs Python code inside a Firecracker microVM. Drop-in replacement for E2BCodeInterpreterTool.
  • IsorunShellTool — runs arbitrary shell commands.

Both tools cache one sandbox per tool instance and reset the auto-destroy timer on every call so the sandbox stays alive across multiple agent turns.

With LangGraph

from isorun.integrations.langchain import IsorunCodeInterpreterTool
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent


with IsorunCodeInterpreterTool() as tool:
    agent = create_react_agent(
        ChatOpenAI(model="gpt-4o"),
        tools=[tool],
    )
    result = agent.invoke({
        "messages": [("user", "Compute the first 100 prime numbers and print them.")]
    })
    print(result["messages"][-1].content)

With LangChain agents (legacy AgentExecutor)

from isorun.integrations.langchain import IsorunCodeInterpreterTool, IsorunShellTool
from langchain.agents import AgentExecutor, create_openai_tools_agent
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI


code  = IsorunCodeInterpreterTool()
shell = IsorunShellTool()
tools = [code, shell]


prompt = ChatPromptTemplate.from_messages([
    ("system", "You write and run code to solve user tasks."),
    ("user", "{input}"),
    ("placeholder", "{agent_scratchpad}"),
])
agent = create_openai_tools_agent(ChatOpenAI(model="gpt-4o"), tools, prompt)
executor = AgentExecutor(agent=agent, tools=tools)


try:
    result = executor.invoke({"input": "Install requests, fetch example.com, return the title"})
    print(result["output"])
finally:
    code.close()
    shell.close()

Drop-in for users coming from e2b

The package exports E2BCodeInterpreterTool as an alias of IsorunCodeInterpreterTool. Existing e2b code keeps working with one import change:

# Before
from langchain_e2b import E2BCodeInterpreterTool


# After — just change the package
from isorun.integrations.langchain import E2BCodeInterpreterTool

Custom image

Pass any OCI image to the tool constructor. First boot pulls + builds a golden snapshot (~30 s); every subsequent invocation is the standard 27 ms cold boot.

tool = IsorunCodeInterpreterTool(
    image="python:3.12",     # default
    sandbox_timeout=600,     # auto-destroy after 10 min idle
    allow=["api.openai.com"],# egress allow list
)

What you get for free

  • KVM hardware isolation (the agent’s code can’t escape the VM)
  • Per-second billing — pay only for the seconds the tool actually used
  • 27 ms cold start vs e2b’s ~150 ms
  • sb.fork(), sb.shell(), sb.url(port), sb.hibernate() etc. are all available on the underlying IsorunCodeInterpreterTool._sandbox if you need them