Docker Sandbox

[methanoya]

See Docker Sandbox setup to provide a seamless and secure way to run the project. PR #445

[MadsLorentzen]

Thanks for opening this alongside the PR. I have replied on #445 with the full reasoning - the short version: environment harnesses live in forks first, the same answer the Nix proposal got in #243, and the way to make yours findable is a post in #78 once the .sbx/ directory sits in your fork. Happy to keep this thread for questions about that, or for anyone comparing sandbox approaches.

[methanoya]

The SBX is the perfect way to run any AI agents securely. Most of the points in the artificial analysis are invalid:

• Docker Sandbox doesn't stall after the first start - it just can have a different release cadence than Claude, for example. So Claude can be outdated in the official image the day Anthropic released the new CLI version. So, the sandbox catches up from time to time, just like any local setup.
• The single source of truth issue can be solved only if you provide the holistic installation script. That script can be hooked up in environments other than the manual one. .github/workflows/ci.yml cannot be reused for the local setup in its current shape.
• Running any native AI installation is reckless, and the community is going to realize that. The SBX is the easiest way to solve this problem since it provides a stable, reproducible environment and eliminates security threats.