OpenAI Agents

pip install isorun[openai]

isorun_tools() returns a tuple of (tool_list, cleanup_fn). Pass the tool list to Agent(tools=...) and call the cleanup function when the agent loop is done. The tools share a single cached sandbox so multi-turn state (cwd, env vars, files written by previous calls) persists across tool invocations.

Example agent

from agents import Agent, Runner
from isorun.integrations.openai import isorun_tools

tools, cleanup = isorun_tools(image="python:3.12")

try:
    coder = Agent(
        name="coder",
        instructions=(
            "You are a senior Python developer. Use the code_interpreter "
            "tool to write and run Python that solves the user's task."
        ),
        tools=tools,
    )
    result = Runner.run_sync(coder, "Compute 2**100 and explain the result.")
    print(result.final_output)
finally:
    cleanup()

What gets exposed

Tool	What it does
code_interpreter(code: str)	Executes Python in a Firecracker microVM with internet access, pip, and a full Python environment. Returns stdout/stderr.
shell(command: str)	Executes a shell command in the same sandbox. Returns stdout/stderr.

Both tools share the same underlying isorun sandbox, so state created by shell (e.g. git clone, pip install) is visible to subsequent code_interpreter calls and vice versa.

Custom config

tools, cleanup = isorun_tools(
    image="python:3.12",      # any OCI image
    api_key="isorun_live_<region>_<id>_<hmac>",    # default: ISORUN_API_KEY env var
    api_url="https://api.isorun.ai",
    sandbox_timeout=900,      # auto-destroy after 15 min idle
)