What it is
An open-source runtime for AI agents in isolated cloud sandboxes: CLI, HTTP API, TypeScript SDK, and chat bots for Slack, Discord and Telegram. Pluggable backends: E2B, Vercel Sandbox, Docker, Cloudflare Containers.
The idea
Traditional sandboxing is code-driven: you write code, then sandbox the code. Sandcaster is data-driven: the agent is a package (skills, behavior, everything) described by a single config file. Load the package and the framework knows how to run it, with well-defined inputs and outputs. Not revolutionary, but a different way to think about shipping agents: agents as artifacts you distribute, not deployments you build.
Why I built it
I’d seen the idea done well in Python and wanted to know if I could build my own take in TypeScript, on my stack, my way. It’s an exploration, not a business. I say that plainly because pretending otherwise is how side projects rot. It’s also the direct descendant of the sandboxing problems I worked on during the agents pivot at Access.
What I learned
Rewriting something that exists is underrated as a learning instrument. You inherit the problem definition and spend all your energy on the solution shape. And “agent as data” keeps looking righter the more the ecosystem standardizes around protocols instead of frameworks.