Anthropic has launched self-hosted sandboxes and MCP tunnels for Claude Managed Agents, enhancing the computational environment for AI agents. The agent loop, responsible for task orchestration and error recovery, remains on Anthropic's cloud, while tool calls and code execution are now handled on enterprise infrastructure or managed sandboxes like Cloudflare and Vercel. This allows enterprises to allocate resources independently and utilize on-premises computing for complex tasks. The new MCP tunnel feature provides an encrypted connection between agents and internal services, enabling access to internal databases and private APIs without opening inbound firewalls. Sensitive data remains within the enterprise's security perimeter. Self-hosted sandboxes are currently in public beta, and MCP tunnels are available for research preview applications.