Sandbox CLI Agents with Vibekit (github.com)

Superagent Technologies released VibeKit, a CLI and SDK that runs coding agents (Claude Code, Gemini, Codex, Grok, OpenCode and others) inside local, isolated Docker sandboxes with built‑in redaction and observability. Install with npm install -g vibekit and run agents like vibekit claude; agent outputs execute in ephemeral containers so your host environment is protected. VibeKit auto-scrubs secrets and API keys from completions, provides real‑time logs, traces and metrics for full visibility, and works entirely offline — no cloud dependency required. SDK and auth packages (npm install @vibe-kit/sdk, @vibe-kit/auth) let you embed sandboxed agent execution and MAX subscription auth into applications. For the AI/ML community, VibeKit addresses a core risk of deploying or experimenting with code‑generating models: executing untrusted/generated code safely while retaining auditability and privacy. Its Docker isolation plus automatic redaction helps teams build secure pipelines, reproduce experiments, and comply with data policies without shipping data to third parties. Observability makes debugging and monitoring agent behavior practical in research and production, and the universal agent support means you can standardize a single secure runtime across multiple model CLIs. Open‑source (MIT) and contribution friendly, VibeKit is positioned as a practical tool for teams that need to run LLM code agents confidently on local infrastructure.