It's built as a base for quickly prototyping domain-specific custom agents. You can fork the repo, point Claude Code at it, and quickly get to a working and usable prototype.

I built this after collaborating with a team that's building an agent for biology. If they used something like this as a base when they started, I think it would have sped things up at the start.

This is yet another agent sandboxing project, but this one is interoperable with Claude Code's built-in session management. It does this by using the three arguments that I learned about from trying out claude.ai/code: --replay-user-messages, --sdk-url, and --resume. The --sdk-url arg enables live interaction over a WebSocket, and the --resume arg pulls session history from a URL.

This starter includes persistent sessions that are interoperable with what Claude Code expects, event streaming over a WebSocket, a web app, and spawning sandboxes through Modal or a Docker host.

Long-lived credentials stay outside of the sandbox; it has a basic credential-injecting proxy for the Anthropic API and GitHub. The proxy limits git pushes to a session-scoped branch.

There's still a gap between forking this app and charging users for a custom agent, but I think it's a somewhat sane set of choices for quickly prototyping.

(I used to work at Anthropic, but this isn't affiliated with them and was built solely with public information.)