LLM Sandbox

LLM Sandbox runs locally from a Python package published to PyPI: llm-sandbox version 0.3.43. It speaks MCP over stdio, so the client starts it as a program and talks to it through standard input and output. It needs Python; clients usually start it with uvx (from uv) or after pip install — the usual command is uvx llm-sandbox. It reads these environment variables: BACKEND, DOCKER_HOST, KUBECONFIG, NAMESPACE, COMMIT_CONTAINER, KEEP_TEMPLATE, SANDBOX_NETWORK_MODE, SANDBOX_READ_ONLY, SANDBOX_CAP_DROP, SANDBOX_SECURITY_OPT, SANDBOX_MEMORY and SANDBOX_CPUS; set them in the client's configuration for this server. Required arguments: --from.

In the mcpServers JSON format that many desktop and editor MCP clients read, the entry looks like this (placeholders in angle brackets):

{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ],
      "env": {
        "BACKEND": "<value>",
        "DOCKER_HOST": "<value>",
        "KUBECONFIG": "<value>",
        "NAMESPACE": "<value>",
        "COMMIT_CONTAINER": "<value>",
        "KEEP_TEMPLATE": "<value>",
        "SANDBOX_NETWORK_MODE": "<value>",
        "SANDBOX_READ_ONLY": "<value>",
        "SANDBOX_CAP_DROP": "<value>",
        "SANDBOX_SECURITY_OPT": "<value>",
        "SANDBOX_MEMORY": "<value>",
        "SANDBOX_CPUS": "<value>"
      }
    }
  }
}

Derived from the registry entry, not tested here. What the server does, and on what terms, is set by its publisher; check its repository or website before giving it access to your accounts or files. How to add an MCP server to an assistant · Before you connect