Open-source infrastructure for running Claude Agent SDK in Modal's secure sandboxes. Features two execution patterns (ephemeral for batch jobs, persistent for low-latency APIs), HTTP endpoints with streaming, MCP tool integration, and persistent storage. MIT licensed. Perfect for building coding assistants, research agents, or autonomous AI workflows.

I built Agent Sandbox Starter because I needed a production-ready way to run the Claude Agent SDK in isolated environments.

The problem: Runnig autonomous AI agents safely requires proper sandboxing, but setting up the infrastructure is complex an time-consuming.

The solution: This open-source template gives you everything out of the box - Modal handles the secure sandboxing, while you focus on building your agent logic.

What makes it different:
- Two execution patterns to fit your use case (ephemeral for batch jobs, persistent for APIs)
- Built-in HTTP endpoints with streaming responses
- MCP tool integration for extending agent capabilities
- Persistent storage that survives sandbox restarts

Would love to hear your feedback! What features would you like to see added?

Very strong technically, but it may be a bit dense for non-experts. Consider leading with the main benefit before listing features.

Agent Sandbox Starter

Run Claude Agent SDK in sandboxed environments

Run Claude Agent SDK in sandboxed environments