Daytona

Daytona is a managed sandbox infrastructure built specifically for AI agents that need to execute untrusted, model-generated code. It spins up ephemeral Linux, Windows, or macOS environments in under 90 milliseconds, exposes programmatic APIs for process execution, file system access, Git, and LSP, and lets sandboxes stay stateful so long-running agent workflows can pick up where they left off. Snapshots, shared volumes, and multi-region deployment round out the runtime story.

The target user is anyone building coding agents, code interpreters, eval harnesses, data-analysis tools, RL training loops, or computer-use agents who would otherwise hand-roll Firecracker or gVisor. Pricing is pay-per-second compute starting around $0.000014/sec with $200 in free credit, plus surcharges for GPUs (H100, RTX PRO 6000) and Windows. SOC 2, HIPAA, and GDPR are covered, and you can bring your own cloud for customer-managed compute.

The core is open-source so you can audit the isolation model, and there's a REST API plus SDKs in Python and TypeScript. It sits in the same competitive lane as E2B, Modal sandboxes, and CodeSandbox SDK, with the differentiator being explicit support for computer-use agents (virtual desktops with SSH and browser VS Code) alongside headless code execution.