Sandbox Ranking Battle

If an AI agent is going to execute code, browse the web, access files, and use tools, the author argues it's better to use a sandbox rather than running it directly on your local machine, and ranked open-source options based on their own criteria.

The comparison criteria were easy start, snapshotting, fork/clone, pause/resume, Linux + macOS support, and computer-use agent/full desktop support.

• #1 SmolVM: Easy to set up locally, supports Linux/macOS, snapshotting, pause/resume, and persistent sandbox workflows, covering everything from browser sessions to full desktop computer-use flows.
• #2 OpenSandbox: Strong for GUI agents and desktop/VNC workflows, but heavier and more platform-oriented, so it was less preferred on DX grounds.
• #3 Microsandbox: Its local-first, lightweight, simple developer experience is a strength, but snapshotting/clone semantics and full desktop support were judged to be still less certain.
• #4 E2B: Has a low barrier to entry and offers pause/resume and desktop sandboxes, but the author ranked it lower, preferring local, open infrastructure and stronger control.

The core message is that sandbox discussions are effectively mixing up different product categories under one name. Some are simple code runners, some are full agent sandboxes, and some are platforms that include browser, desktop, and computer-use support.

So the author concludes that what really matters isn't "can it run code" but operational characteristics like maintaining state, recovering after an interrupt, browser/desktop use, and fast fork/clone from saved state.