“1. E2B — probably the closest match to what you're describing.”
“Each sandbox is a Firecracker microVM with its own kernel.”
“You can pause it and later resume with memory and filesystem state intact, and you can destroy it when finished.”
“It also has network/egress controls, which are particularly important when executing hostile code.”
“E2B explicitly restores memory and filesystem state after pause/resume.”
“2. Daytona — excellent if the sandbox is more like a temporary development machine.”
“Daytona's VM sandboxes use Firecracker and provide a full Linux environment, including shell, filesystem, processes and networking.”
“Its API is designed for creating, stopping and deleting sandboxes, and it supports environments that persist much longer than typical short-lived code runners.”
“3. Vercel Sandbox — attractive if you're building a web product on Vercel.”
“It uses Firecracker microVMs and gives your application programmatic control over sandbox execution.”
“Persistent sandboxes retain filesystem state, while disposable sandboxes are appropriate for one-shot execution.”
“4. Modal — particularly good if generated code needs GPUs or substantial compute.”
“Modal creates isolated disposable sandboxes and explicitly documents the pattern of creating one, executing generated code, and terminating it afterward.”
“Modal offers gVisor/VM options.”
“5. Cloudflare Sandboxes — compelling if your application already lives in Workers.”
“Cloudflare explicitly positions these for untrusted or generated code.”
“A sandbox can run a Linux container, and the Worker controls what application data and network access the code receives.”
“Fly.io Sprites | Firecracker microVM | Warm suspension preserves processes | Yes | Persistent Linux workspaces”
“Prized | Firecracker microVM | Yes — memory-preserving pause + fork | Yes | E2B-like sandbox API”
“If you need to continue an agent session exactly where it left off—including in-memory state—E2B and Prized are especially interesting.”
