zanetworker/sandboxing-anywhere

Agent sandboxing experiments across providers: OpenShell, OpenAI Containers API, Anthropic self-hosted sandboxes, and more. Each directory is a self-contained integration example.

★ 1Forks 1PythonGitHub ↗Compare

README

Sandboxing Anywhere

Agent sandboxing experiments across providers and entrypoints. Each directory is a self-contained integration example with its own dependencies and instructions.

OpenShell + OpenAI Agents SDK is the first working integration. The upstream PR adding OpenShell as a sandbox provider to the OpenAI Agents Python SDK: openai/openai-agents-python#3469

Experiments

Directory What Sandbox Provider Agent Framework
openshell-agents-sdk/ OpenShell + OpenAI Agents SDK NVIDIA OpenShell OpenAI Agents Python SDK
openshell-agents-sdk-self-hosted/ OpenShell + self-hosted vLLM (fully air-gapped) NVIDIA OpenShell OpenAI Agents SDK + vLLM on GPU cluster
openshell-ogx/ OpenShell + OGX (Llama Stack) NVIDIA OpenShell OGX Containers API + Responses API (PR)
anthropic-self-hosted/ OpenShell + Anthropic Self-Hosted Sandboxes NVIDIA OpenShell Anthropic ant CLI (Managed Agents, self-hosted)
supervisor-pattern/ Multi-agent supervisor: brain decomposes, each agent reasons with vLLM inside its own OpenShell sandbox with per-agent network policy NVIDIA OpenShell Claude Managed Agents (brain) + vLLM/Qwen3-8B (sandboxed local agents)

Related PRs and Issues

Link Status Description
ogx-ai/ogx#5853 Open PR Containers API + Docker provider + remote::openshell provider
ogx-ai/ogx#5852 Open Issue Containers API feature request
openai/openai-agents-python#3468 Open Issue Feature request: add OpenShell sandbox provider
openai/openai-agents-python#3469 Open PR OpenShell sandbox extension for the OpenAI Agents SDK
RHAIRFE-1538 Stakeholder Review Sandboxed Containers API and Shell Tool for Agent Code Execution

Planned

  • openai-containers-api/ -- OpenAI Responses API with Containers API (hosted sandboxes)
  • openshell-mcp/ -- OpenShell sandboxes exposed as MCP tools
  • openshell-llama-stack/ -- OpenShell + Llama Stack agents (native API, not OpenAI-compat)

Structure

Each experiment directory contains:

<experiment>/
  README.md          # Setup instructions and usage
  pyproject.toml     # Dependencies (or requirements.txt)
  src/               # Source code
  examples/          # Runnable examples

Prerequisites

All experiments assume you have the relevant sandbox provider running. See each experiment's README for specific setup instructions.

Contributors

zanetworker

Issues