Perplexity open-sources its Firecracker sandbox machinery

What it does

SPACE (Sandboxed Platform for Agentic Code Execution) creates isolated Linux environments in Firecracker microVMs and lets your application run commands in them, stream output, and manage the sandbox lifecycle. You define a template — the starting environment — and sandboxes are instances of it, each with its own VM and writable filesystem. Your app decides what to run and what to keep; SPACE handles the VMs, processes, and the data-plane connection carrying command output.

The interesting bit

The boring part is the value: lifecycle operations are asynchronous with pollable operation IDs, command output flows over a separate connection with session-scoped attachment tokens, and sandboxes can be paused, suspended, and woken later. There’s also an E2B compatibility wrapper, so code written against the E2B SDK can talk to SPACE — a pragmatic migration path, not a rewrite.

Key highlights

• Runs each sandbox in its own Firecracker microVM with a writable filesystem; multiple commands share one sandbox
• HTTP API plus server-sent events — no Rust library to embed; the docs include a stdlib-only Python example
• Warm pools keep booted sandboxes ready so creates skip a cold boot; base snapshots restore sandboxes on a single node
• E2B compatibility wrapper with a tested E2B Python SDK path
• Kubernetes is the documented deployment, but the architecture guide covers running without it

Caveats

• The documented install path is a local Kind cluster — this is operator-grade software, not a drop-in library
• Schema fields marked incubating are explicitly outside the supported contract, so pin your expectations
• E2B compatibility has documented behavior limitations, not full parity

Verdict

Worth a look if you’re building an agent platform and need real VM isolation with pause/suspend semantics. If you just want to run untrusted snippets occasionally, a lighter sandbox layer will cost you less operational love.