Th0rgal/sandboxed.sh

Safe runtime for autonomous on-chain AI agents: isolated sandboxes, Library skills, encrypted secrets, and OKX read-only security checks.

What it solves

sandboxed.sh is a self-hosted cloud orchestrator designed to run AI coding agents safely and autonomously. It solves the problem of giving powerful agents (like Claude Code or Gemini) the ability to execute code, manage files, and run tests in an isolated environment without risking the host machine's security or stability.

How it works

The system creates isolated Linux workspaces using systemd-nspawn or Docker containers. It orchestrates multiple agent runtimes—including Claude Code, OpenCode, Codex, Gemini, and Grok—within these containers. Users can manage these agents through a web dashboard or an iOS app, monitoring their progress in real-time via streaming logs and resource graphs. It also includes a Git-backed library for versioning skills, tools, and rules, and a model routing system to handle provider fallbacks and rate limits.

Who it’s for

Developers and AI engineers who want to deploy autonomous coding agents that can handle entire development cycles (e.g., fixing GitHub issues and opening PRs) or long-running operations (like fine-tuning models) while keeping their data local and secure.

Highlights

  • Multi-Runtime Support: Compatible with Claude Code, OpenCode, Codex, Gemini, and Grok.
  • Isolated Workspaces: Uses containerized Linux environments to scope bash commands and file operations.
  • Mission Control: Remote monitoring and management of agents with real-time streaming.
  • Git-backed Library: Version-controlled skills, tools, and MCPs.
  • Model Routing: Provider fallback chains with health checks and rate-limit handling.
  • Multi-platform: Includes a Next.js web dashboard and a SwiftUI iOS app with Picture-in-Picture support.