Open-source barebox sandbox: r/w/x permissions let AI agents run with 1-second startup
Tight_Heron1730 · reddit · 2026-11-05
The author released barebox, an open-source (Apache-2.0) sandbox addressing the dilemma of giving AI agents real computer access: approving every step until you stop reading, or heavy isolation most people skip.
Core design: every action gets one letter, like file permissions —
- r: read/lookup, just happens;
- w: undoable changes kept on an undo layer you keep or discard;
- x: irreversible actions (paying, sending, deleting) wait for your yes.
Three modes: ask (confirm before any change) / auto (default, only irreversible needs you) / trust (runs freely). Payments ask in every mode.
Lightweight: no setup, no root, no Docker image, 8 MB each, up to 40 side-by-side on one laptop. In the demo, Meta's muse model books a trip autonomously, stopping only for 8 irreversible actions and never seeing the real payment key. Honest limit: it's a guardrail against mistakes and tricked agents, not a fortress.
More from coding & agent
- Levels decompiles Battlefield 1942 with Claude Code, ports it to browser with working multiplayer — huangyun_122 · 2026-11-05
- Kubernetes promotes AI Conformance to permanent SIG Architecture subproject — TerryTangYuan · 2026-11-05
- quiet.md goes open source: Rust+GPUI editor, Feishu-compatible, ~5ms frames on 1M-char docs — dotey · 2026-11-05
- A background agent that auto-structures messy scientist Excel data, day 17 of agent series — nlarusstone · 2026-11-05
- Code mode isn't about MCP: it's about whether tool results go to context or program variables — ritakozlov · 2026-11-05
- Agent Arena: frontier models show 55-minute gap on long-horizon agent tasks — arena · 2026-11-05