Show HN: Secure Code Execution for AI (github.com)

ERA is an open-source runtime for securely executing arbitrary code and long-running AI agents with persistent storage and sandboxed VMs. The system splits a Go-based era-agent (VM orchestration, lifecycle management, multi-language runtimes) from a Cloudflare Worker deployment layer that routes requests, manages sessions with Durable Objects, and stores files and packages in R2. It supports Python, Node.js/TypeScript, Go and Deno, automatic npm/pip package installation, and can run standalone (Docker, K8s, bare metal) or globally on Cloudflare’s edge—Cloudflare builds your Docker image so no external registry is required. Primary HTTP APIs include ephemeral execution (POST /api/execute), persistent sessions (POST/GET/DELETE /api/sessions), and low-level VM controls; local dev tooling and a CLI are provided. For the AI/ML community this matters because it gives a reproducible, multi-tenant sandbox for running model-generated code, chaining agent workflows, and persisting intermediate state safely—useful for model-in-the-loop data pipelines, agent orchestration, CI/CD test runs, and educational sandboxes. Key technical implications: isolated per-VM execution, persistent storage of packages/files via R2, session metadata in Durable Objects, and extensibility across deployment targets. Getting started requires Node 18+, Docker Desktop for local dev, and optionally Go 1.21+ to build era-agent; docs and examples are included for quick deployment and integration.