E2B

How E2B works

E2B runs each session inside its own Firecracker microVM, the same lightweight hardware virtualization used for serverless functions, so agent-generated code executes behind a real kernel boundary rather than in a shared-kernel container. Cold start is around 150ms. Sessions can run up to 24 hours, and you can pause one to keep its memory state or take a filesystem snapshot and resume later. Billing is per second. The TypeScript and Python SDKs are the main interface, and E2B is an official OpenAI Agents SDK provider, which makes it a natural pick when isolation for untrusted code is the priority.

Pricing in practice & watch-outs

Billing is per second at about $0.05 per vCPU-hour, so the bill largely follows active compute. The paid Pro plan starts at $150 a month, a real jump from the $100 one-time Hobby credit once you move past prototyping. The other limit is hardware: there is no GPU inside the sandbox, so accelerated inference has to run on a separate service and reach the sandbox over the network. For CPU-bound agent code execution the pricing is easy to reason about. For GPU work, E2B is not the tool.