Anthropic has split your AI agent in two. The company just released self-hosted sandboxes and MCP tunnels for Claude Managed Agents, separating the brain from the body.
The agent loop — task orchestration, context management, and error recovery — stays on Anthropic's cloud. But all tool calls and code execution now run on the enterprise's own infrastructure, or on managed sandboxes from Cloudflare, Daytona, Modal, and Vercel.
Companies can assign their own CPU, memory, and runtime images. Agents that need to build something over a long period or generate images can tap directly into local compute.
The MCP tunnels feature creates an encrypted channel through a lightweight, one-way gateway deployed inside the corporate network. No inbound firewall rules needed. Claude agents can call internal databases, private APIs, and knowledge bases — without sensitive files or credentials ever leaving the enterprise security boundary.
Self-hosted sandboxes are in public beta. MCP tunnels are available as a research preview.