AI Agent Sandboxes
Full Linux computers
for your AI agents
Isolated VMs that start in under a second. Drive them from the SDK or CLI.
- < 1s
- Boot time
- $0.00084
- Per vCPU / minute
- Per second
- Billing
Prepare, run, pause, branch, share.
Put your agent on your tailnet
Each sandbox can join your Tailscale network as its own device. Your agent reaches private services by their MagicDNS names, and your tailnet reaches only the ports you expose. Need a fixed public IP? Route its internet traffic through your own exit node.
Read the Tailscale guideYour custom image, running in under a second.
We own the compute, storage, and network, so the whole startup path is built around your image.
Create a custom imageOwned end-to-end.
Compute, Ceph storage, switches, power, and uplink are ours.
Mount the disk. Skip the copy.
Your disk mounts straight from Ceph. Nothing is copied before boot.
Drive sandboxes from your terminal or code.
Create a VM, run a command, read the result. TypeScript, Go, Python, or the CLI.
- 01Create a VM
createAndWait() - 02Run a command
session.exec() - 03Read the result
result.stdout
Pay for what runs. Upgrade your workspace to raise limits.
Billed per second. Per-minute rate shown for reference. First 5 GiB of disk free.
View pricing| Compute | Per minute | Per second |
|---|---|---|
| CPU/ vCPU | $0.00084 | $0.000014 |
| Memory/ GiB | $0.00027 | $0.0000045 |
| Disk/ GiBFIRST 5 GiB FREE | $0.0000018 | $0.00000003 |
Enterprise-grade security.
Each sandbox is a disposable, hardware-isolated VM.
Hardware-isolated VMs
Its own kernel, not a shared-kernel container.
Wiped on teardown
Disk and memory are gone when the session ends.
Secrets never cross boundaries
Credentials are scoped to one session.
Encrypted in transit
Tunneled, authenticated connections. No public IPs.
Don't believe us. Believe them.
Questions? Answered.
Have a question we haven't answered? Reach out to hello@luxor.ai
A sandbox is an isolated, fully managed compute environment where developers and AI coding agents can write, run, and test code without affecting production systems. Think of it as a disposable virtual machine spun up on demand.
Luxor Agent Cloud provisions sandboxes from custom images in under a second, including your tools and dependencies. Sandbox disks are mounted directly from our dedicated Ceph storage cluster instead of provisioned from the compute node’s local drive. This keeps provisioning fast for your own environment as well as our base images.
Full VM sandboxes are a great fit for agents that need to run tests, compile code, or interact with Docker. Every agent task gets its own isolated workspace, so agents can work with root access without ever touching your host machine.
Yes. Full VM-based sandboxes support Docker, system packages, and background services. This makes them ideal for integration tests and workflows that mirror a CI/CD environment exactly.
Yes. Define a typed template with your repository, dependencies, and build steps, then build it into a reusable image. You can also snapshot a running sandbox and start new sessions from that saved state. Templates do not currently accept Dockerfiles.
Pausing stops compute charges and preserves memory and durable files so you can resume later. Active connections close and /tmp is cleared; keep durable files under /home/tenki or on a volume. A non-sticky session is paused automatically when it reaches its maximum duration. Terminating ends the session permanently; saved snapshots and persistent volumes are managed separately.
Usage is metered per second across CPU, memory, and disk. Pausing a sandbox stops compute charges. Sandboxes are not paused for inactivity: a running sandbox accrues compute charges until you pause or terminate it, or until a non-sticky session reaches its maximum duration and is paused. The first 5 GiB of disk is included, and resource limits scale with your workspace plan.
Spin up your first environment in seconds with the TypeScript, Go, or Python SDK, or the CLI.











