LLM Sandbox

LLM Sandbox is a local MCP server: Securely run LLM-generated code in isolated containers across 7 languages and 3 container backends. Install it from pypi as llm-sandbox.

Claude CodeCodexAntigravity

commandprompt

Terminal

claude mcp add llm-sandbox -e BACKEND=... -e DOCKER_HOST=... -e KUBECONFIG=... -e NAMESPACE=... -e COMMIT_CONTAINER=... -e KEEP_TEMPLATE=... -e SANDBOX_NETWORK_MODE=... -e SANDBOX_READ_ONLY=... -e SANDBOX_CAP_DROP=... -e SANDBOX_SECURITY_OPT=... -e SANDBOX_MEMORY=... -e SANDBOX_CPUS=... -- uvx llm-sandbox

Run in your project folder. Add --scope user to use it everywhere.

Add to claude_desktop_config.json or .cursor/mcp.json.

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

Fill in the environment variables with your own credentials.

• Runtime: Python with uv (runs with uvx).
• MCP client: Claude Code, Codex, Antigravity, Claude Desktop, Cursor or any other client that speaks MCP over stdio.
• 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, SANDBOX_CPUS