Spawn secure, disposable Firecracker microVMs for your AI agents in one API call. Run code, drive a browser, open an IDE — then vanish without a trace.
An AI agent receives a task and spawns sandboxes through the spawn API: it runs code in an isolated Firecracker microVM, drives a headless browser, opens a cloud IDE, or fans out a swarm of sandboxes in parallel, then answers and destroys or pauses them.
1963 — an IBM 7090 filled this room at NASA Ames. Today, every agent gets its own computer in one API call.
Photos: NASA
maya · just now
@spawn add Apple Pay to checkout. Keep the old flow behind a flag, and cover it with tests.
Add Apple Pay to checkout
✓ All checks have passed · ready for review
Payment session endpoint
booting…
Apple Pay button
booting…
apple_pay feature flag
booting…
Checkout e2e tests
booting…
Claude Monet, The Houses of Parliament, Sunset, 1903
How it works
spawn.create()Every sandbox is its own Firecracker microVM on a KVM host. This is the path a request takes — from your API key to a running machine, and back to nothing.
The control plane checks your key, the size you asked for and your limits. The orchestrator samples a few healthy workers and picks the one with the most headroom — and that worker reserves CPU, memory, disk and a network slot before anything boots.
The worker restores a prebuilt Firecracker template — VM state, memory and a copy-on-write disk — inside its own cgroup and network namespace. The API returns the moment the guest agent inside the VM answers.
Run and exec skip the orchestrator: the control plane talks straight to your sandbox’s worker, which keeps a persistent vsock connection to the guest. Python and Node stay warm between calls, so state carries over.
Attach a live terminal over WebSocket, open VS Code in the browser, or expose any port through a signed preview link. Every link is single-use or expiring, and private by default.
Pause snapshots memory and VM state, uploads only the disk chunks that changed, and frees the CPU and RAM. Resume restores it on the same worker. Destroy tears down the VM and wipes its disk — only the audit record stays.
Features
One small API on top of hardware-isolated microVMs. No containers to harden, no infra to babysit.
Create a sandbox, run code, read the output — over plain HTTP with an API key. SDKs land at launch.
Keep a sandbox alive between calls. Files, variables and installed packages survive across steps.
Python, Node, Bash and Go out of the box. A full Linux VM, not a restricted interpreter.
Feed in CSVs, JSON or PDFs, let your agent transform them, and pull the results back out.
Connect a repo through our GitHub App, pin a commit, and run that exact code in a fresh sandbox.
Open an interactive shell into any running sandbox from the dashboard. See exactly what your agent saw.
Every execution is logged and metered per second — see what ran and exactly what it cost.
Firecracker microVMs — the technology behind AWS Lambda — give every sandbox its own kernel.
Hardware isolation
Each sandbox is a Firecracker microVM with its own kernel, memory and disk. Nothing is shared while it runs, and nothing is left behind when it’s gone. Isolation is the foundation. Everything we build next starts here.
Private beta · $20 in free credit · no card required
spawn-review bot
SQL injection.
qgoes straight into the query string. Pass it as a parameter: