Railway Sandboxes now ship with Claude Code, OpenAI Codex, OpenCode, and Pi pre-installed in the default image, eliminating manual agent setup friction. The post walks through the recommended workflow: build a reusable template, create a sandbox, configure it, checkpoint the state, then create or fork sandboxes from that checkpoint for parallel agent workloads. Sandboxes can join a Railway environment's private network, giving agents access to Postgres, Redis, and other project services. The TypeScript SDK enables programmatic sandbox creation, long-running detachable agent sessions, and forking for parallel agent runs. The post also clarifies the differences between templates (repeatable bases), checkpoints (named disk snapshots after expensive setup), and forks (live clones for branching work).
