Sandcastle (AI)

Canonical version: Sandcastle (AI).

Sandcastle is a TypeScript library by Matt Pocock for orchestrating AI coding agents in isolated sandboxes. You invoke an agent with a single sandcastle.run() call; the library runs it in a sandbox on a configurable branch, then merges the resulting commits back into your repo. It targets parallelizing multiple AFK agents, building review pipelines, or orchestrating your own agents.

Key aspects

• Provider-agnostic: Docker, Podman, Vercel, or custom providers. Bind-mount providers (like Docker) mount worktrees directly; isolated providers sync via copyIn / copyFileOut
• Branch strategies: head (direct write), merge-to-head (temporary branch), or explicit branch names
• Session capture and resume for continued agent conversations
• Structured output via Standard Schema validators
• Prompt system with dynamic context (shell expressions, template substitution) and before/after lifecycle hooks
• Multi-iteration support with completion signals and idle timeouts

Agent support

Claude Code, Codex, Pi, Cursor, OpenCode, and Copilot, each configurable (model, reasoning effort, permission modes). MIT licensed; TypeScript (tsup + vitest).

References

• https://github.com/mattpocock/sandcastle

Related

• Matt Pocock
• AI Agents
• Agentic Engineering
• Docker Sandboxes
• OpenSandbox
• Vercel Sandboxes
• Cloudflare Sandbox SDK
• microVM
• Claude Code
• TypeScript
• MIT License

About Sébastien

Ready to get to the next level?