The Execution Engine for AI Agents
Execute code, automate desktops, boot from Docker images, expose ports, clone live state, or SSH in directly. Each VM gets a dedicated kernel with sub-200ms cold start and deny-by-default networking.
Code Execution
Execute untrusted code in isolated microVMs
Run untrusted or model-generated code in Firecracker microVMs. Each call to client.execute() gets a dedicated kernel, filesystem, and network stack.
<200ms
P95 cold start
<10ms
Warm reuse
Learn more →
Computer Use
Full Linux desktop for agent automation
Give agents a full Linux desktop with browser, terminal, and sudo. Proxy mouse/keyboard actions through the API. Watch live via noVNC viewer.
1920x1080
Default viewport
All HTTP methods
Proxied
Learn more →
Run Any OCI
Boot VMs from any OCI or Docker image
Pass any Docker Hub or private registry image to snapshots.create(). The snapshot pre-installs packages and clones repos. New VMs boot from it in <500ms.
<500ms
Snapshot boot
3 GB
Max disk
Learn more →
Shares & Deploy
Expose VM ports as public URLs or custom domains
Expose any VM port as a public URL with shares.create(). Toggle visibility, attach custom domains with DNS verification. Works with any app.
1 API call
Port to URL
Public / private
Toggle
Learn more →
Live VM Clone
Clone running VMs for parallel eval fan-out
vms.clone() copies a running VM's full state — disk, memory, processes — into a new VM. Clone once or fan out to many for parallel evals.
Full state
Disk + memory
Source unchanged
Non-destructive
Learn more →
SSH Workflows
SSH into any VM with account-level keys
SSH directly into any VM. Keys are account-level — add_ssh_key() once, connect to all VMs. VM response includes ssh_host, ssh_user, ssh_port.
Account-level
Key scope
ssh_host + ssh_port
Per VM
Learn more →
Get started with InstaVM
Free tier available. No credit card required. First VM provisions in under 200ms.