Cluaiz adds OS sandboxing to its AI agent after three real-machine failures

CommissionOdd3082 · reddit · 2026-11-22

A Cluaiz contributor argues that an agent's allow/ask/deny approval only decides whether a command runs, not what it can touch — approving npm test still let it read /.ssh and /.aws. The open-source AI engine added a second layer: every agent command, MCP server and native plugin now runs inside an OS sandbox (bubblewrap on Linux, sandbox-exec on macOS, AppContainer on Windows), with the project folder writable, the rest of disk read-only, and home/secrets masked. Approved commands keep network; tools declaring no network get none.

Real-machine testing broke it three times:

After fixes: 9/9 checks pass on macOS and Ubuntu runners, 8/8 on an Ubuntu VM. Windows works but is opt-in, since Node.js and per-user Python installs can't start inside AppContainer. The full agent + MCP path in strict mode is unit-tested only and lives on main, not in the 0.1.0 release. Open question: when the sandbox can't start (stock Ubuntu, WSL1, containers), should the agent refuse or warn? They default to warn, refuse in strict mode.

Original post →

More from coding & agent

coding & agent channel →