ModalREPL

Cloud sandbox execution environment

ModalREPL executes Python code in Modal cloud sandboxes, which are ephemeral cloud VMs that run completely isolated from the host machine. Each sandbox is a fresh, isolated environment with its own filesystem, network, and compute resources, providing the highest level of security and isolation available in RLM. The sandbox requests LM calls from the host’s LM Handler when code executes llm_query(). This environment is production-ready and essential for executing untrusted LM-generated code or handling sensitive data.

Prerequisites

uv pip install -e . --extra modal
# Or with regular pip:
# pip install -e ".[modal]"

modal setup

rlm = RLM(
    backend="openai",
    backend_kwargs={"model_name": "gpt-5-mini"},
    environment="modal",
    environment_kwargs={
        "app_name": "my-rlm-app",
        "timeout": 600,
    },
)

Arguments

Argument	Type	Default	Description
app_name	str	"rlm-sandbox"	Modal app name
timeout	int	600	Sandbox timeout in seconds
image	modal.Image	Auto	Custom Modal image
setup_code	str	None	Code to run at initialization
context_payload	str | dict | list	Auto	Initial context (set by RLM)
lm_handler_address	tuple	Auto	Socket address (set by RLM)

How It Works

Modal sandboxes can’t connect directly to the host. Uses HTTP broker pattern:

1. Sandbox starts Flask broker server on port 8080
2. Port exposed via Modal’s encrypted_ports tunnel
3. llm_query() POSTs to local broker, blocks waiting
4. Host polls {tunnel}/pending every 100ms
5. Host forwards requests to LM Handler, POSTs responses back
6. Broker unblocks and returns response

Host polls /pending ────────────────┐
                                    │
┌───────────────────────────────────┼──┐
│ Modal Sandbox                     ▼  │
│  ┌──────────────┐   ┌──────────────┐ │
│  │ Broker Flask │◄─►│ Code Exec    │ │
│  │  /enqueue    │   │ llm_query()  │ │
│  │  /pending    │   └──────────────┘ │
│  │  /respond    │                    │
│  └──────────────┘                    │
└──────────────────────────────────────┘

Custom Image

You can use your own custom Modal images or update the given image:

import modal

image = modal.Image.debian_slim(python_version="3.11").pip_install(
    "numpy", "pandas", "dill", "requests", "flask"
)

rlm = RLM(
    backend="openai",
    backend_kwargs={"model_name": "gpt-5-mini"},
    environment="modal",
    environment_kwargs={"image": image},
)

Default Image

Includes: numpy, pandas, scipy, sympy, requests, httpx, flask, pyyaml, tqdm, dill