sandboxed.sh

What It Is

sandboxed.sh is an AI coding agent orchestration tool that runs on your own server. It does not “do AI” itself; instead, it manages the runtime environments for Claude Code, OpenCode, Codex, Gemini, Grok, and similar tools, including task launch, streamed progress viewing, container workspaces, and configuration sync. Users can start tasks from a Web Dashboard, iOS App, or CLI, making it suitable for handing off hours-long work such as refactoring, issue fixes, and PR generation to agents that keep running in the background.

Core Capabilities

It mainly addresses three issues: session timeouts, code leaving your infrastructure, and limited customization of project behavior. Each project can run in its own isolated container, preventing dependencies and rules from different Node, Python, and other projects from interfering with one another. Teams can also maintain skills, commands, and rules through a Git repository—for example, requiring specific tests to be run whenever the payments module is modified. Its architecture is closer to an infrastructure layer for AI coding agents than yet another model or IDE.

Deployment and Pricing

The main documentation clearly states that the backend is installed on your own server, requiring both a server and a domain name, with systemd, nginx, and SSL configurable. The Dashboard can be deployed to Vercel, run locally with bun dev, or connected to the server via the iOS App. Pricing for the sandboxed.sh software itself is not disclosed; the only cost guidance given is a reference server cost of around $30/month on Hetzner or DigitalOcean.

Pros and Cons

The main advantage is that code stays on your own machine or infrastructure, making it suitable for healthcare, finance, defense, and other scenarios with strict data-boundary requirements. Long-running tasks also do not depend on a local computer staying online. Container isolation and a rules repository help agents better align with team engineering practices. The downsides are that it requires your own server and deployment skills, so the barrier to entry is higher than ready-to-use tools like Cursor. The article does not clarify whether it is open source or closed source, nor does it cover commercial support, permission management, payment methods, or SLA. Actual coding quality still depends on the underlying agent and model.

Who It’s For and Access from China

It is best suited for individual developers and engineering teams with DevOps capabilities who want to self-host AI coding infrastructure, especially organizations whose code cannot be uploaded to third-party SaaS platforms. Access from mainland China is not discussed in the source text, so it should be considered unknown. In addition, Claude, Gemini, and OpenAI-related APIs may involve network and payment restrictions in China, so deployment should include an assessment of proxies, overseas servers, or local model alternatives.