Coding agent attempts
Prepare a repository once, fork parallel solutions and run each test suite independently.
Sandbox infrastructure for AI agents
Spawn an isolated environment in under 120ms. Fork it for parallel work, destroy it when done, and pay from $0.024/hour.
How warm starts work →Five commands. The whole lifecycle.
The CLI creates box1, forks four linked child boxes, pauses and resumes the parent, then destroys the entire disposable fleet.
See how it worksStart a clean microVM for one agent task or untrusted command.
Branch a prepared sandbox into parallel attempts without repeating setup.
Freeze CPU while preserving the process, memory and filesystem state.
Continue the same workload exactly where the sandbox stopped.
Reclaim compute, network and disposable state when the work ends.
Continuously active scenario: 12,000 hours (50 sandboxes × 240 h), 4 vCPU / 8 GB RAM.
Directional estimate, not a quote. The comparison models continuously active 4 vCPU / 8 GB sandboxes at published May 2026 rates. Plans, minimums, storage, networking and regions differ; active-CPU pricing may produce lower bills for I/O-heavy workloads.
Call the lifecycle from your terminal, control plane or agent runtime—or let Claude operate boxes through MCP.
Use the CLI directly, call the lifecycle over HTTP, bring it into your code with the TypeScript SDK or let Claude operate boxes through MCP.
Explore the interfacesEvery sandbox is a Firecracker microVM with its own network namespace. Kill reclaims the VM, network and disposable state.
Prepare a repository once, fork parallel solutions and run each test suite independently.
Run untrusted commands and collect stdout, stderr and exit status through the API.
Pause while an agent waits for approval, then resume with its process state intact.
Expose an application port and share the resulting sandbox URL.
Every plan includes a monthly pool of hours. Smaller sandboxes consume it more slowly.
No credit card. 24 hours included.
spawn create baseWARM / <120MS